CN115409836B - Metal workpiece detection and cleaning control method and system based on image processing - Google Patents

Metal workpiece detection and cleaning control method and system based on image processing Download PDF

Info

Publication number
CN115409836B
CN115409836B CN202211350654.8A CN202211350654A CN115409836B CN 115409836 B CN115409836 B CN 115409836B CN 202211350654 A CN202211350654 A CN 202211350654A CN 115409836 B CN115409836 B CN 115409836B
Authority
CN
China
Prior art keywords
abnormal
cleaning
production
image data
model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202211350654.8A
Other languages
Chinese (zh)
Other versions
CN115409836A (en
Inventor
熊志红
朱峰
郑思温
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Shishang Electronic Technology Co ltd
Original Assignee
Shenzhen Shishang Electronic Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenzhen Shishang Electronic Technology Co ltd filed Critical Shenzhen Shishang Electronic Technology Co ltd
Priority to CN202211350654.8A priority Critical patent/CN115409836B/en
Publication of CN115409836A publication Critical patent/CN115409836A/en
Application granted granted Critical
Publication of CN115409836B publication Critical patent/CN115409836B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B08—CLEANING
    • B08B—CLEANING IN GENERAL; PREVENTION OF FOULING IN GENERAL
    • B08B13/00—Accessories or details of general applicability for machines or apparatus for cleaning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection

Abstract

The application relates to the technical field of metal workpiece cleaning control, in particular to a metal workpiece detection cleaning control method and system based on image processing. According to the method and the device, the metal workpiece image data of the metal workpiece to be detected are obtained, the decision is made on the metal workpiece image data based on the preset parameter layer information optimization and the selected metal workpiece detection and cleaning decision model, the cleaning decision data of each target workpiece unit in the metal workpiece image data are determined, the metal workpiece to be detected is cleaned and controlled based on the cleaning decision data of each target workpiece unit in the metal workpiece image data, the cleaning decision data of each target workpiece unit in the metal workpiece image data are determined in an image data analysis mode, then the metal workpiece to be detected is cleaned and controlled, and the reliability of a cleaning control process can be improved.

Description

Metal workpiece detection and cleaning control method and system based on image processing
Technical Field
The application relates to the technical field of metal workpiece cleaning control, in particular to a metal workpiece detection cleaning control method and system based on image processing.
Background
In the production and use process of the metal workpiece, the metal workpiece is easily polluted by oil stains and dust, so that a lot of impurities remain on the surface of the metal workpiece, and the metal workpiece needs to be cleaned. However, the cleaning control in the related art is generally performed according to a fixed cleaning control manner and strategy, but the specific residual states of different metal workpieces are different, which results in poor reliability of the cleaning control flow.
Disclosure of Invention
The application provides a metal workpiece detection and cleaning control method and system based on image processing.
In a first aspect, an embodiment of the present application provides a metal workpiece detection and cleaning control method based on image processing, which is applied to a cleaning control system, and includes:
acquiring metal workpiece image data aiming at a metal workpiece to be detected;
determining the cleaning decision data of each target workpiece unit in the metal workpiece image data based on the preset parameter layer information optimization and the selected metal workpiece detection cleaning decision model, wherein the cleaning decision data comprises the cleaning time parameter and the cleaning mode parameter of each target workpiece unit;
and cleaning and controlling the metal workpiece to be detected based on the cleaning decision data of each target workpiece unit in the metal workpiece image data.
For instance, in one possible implementation of the first aspect, the method further comprises:
determining a cleaning control thermodynamic diagram of a corresponding metal workpiece production partition according to a cleaning control data set formed by cleaning control data of each metal workpiece to be detected, wherein the cleaning control thermodynamic diagram is used for representing a distribution diagram formed by cleaning thermodynamic measurement values corresponding to each target workpiece unit in the corresponding metal workpiece production partition;
acquiring a production node sequence corresponding to a target workpiece unit meeting preset conditions in the cleaning control thermodynamic diagram, and analyzing whether production activity data corresponding to each production node in the production node sequence has abnormal production behaviors or not, wherein the preset conditions comprise that a cleaning heat power metric value is larger than a preset heat power metric value;
when abnormal production behaviors exist in the production activity data corresponding to any one production node, outputting the corresponding abnormal production node and the production node corresponding to the abnormal production node to a production service terminal for information prompt;
wherein, the step of analyzing whether the production activity data corresponding to each production node in the production node sequence has abnormal production behavior comprises:
and inputting the production activity data corresponding to each production node in the production node sequence into an abnormal production behavior analysis model for abnormal production behavior analysis, and determining the probability expression of the abnormal production behavior of the production activity data, wherein the abnormal production behavior analysis model is obtained by performing abnormal characteristic AI training on the sample production activity data of the metal workpiece.
For example, in a possible implementation manner of the first aspect, before the step of inputting the production activity data corresponding to each production node in the production node sequence into an abnormal production behavior analysis model for abnormal production behavior analysis and determining an abnormal production behavior probability representation of the production activity data, the method further includes:
obtaining a sequence of exemplar production activity data that includes first exemplar production activity data of the metal workpiece and derived exemplar production activity data of the metal workpiece;
and training a target AI analysis model based on the optimization updating of AI model parameters of the derived example production activity data and the optimization updating of AI model parameters of the first example production activity data, and when the target AI analysis model reaching the parameter optimization output condition reaches the parameter optimization output condition, taking the target AI analysis model reaching the parameter optimization output condition as an abnormal production behavior analysis model, wherein the abnormal production behavior analysis model is used for analyzing an abnormal support probability value of abnormal production behaviors in the production activity data of the metal workpiece.
For instance, in one possible implementation of the first aspect, the example production activity data sequence further includes example abnormal support probability values for abnormal production behavior in the first example production activity data;
the training of the subject AI analytical model based on the tuning of AI model parameters for the derived example production activity data and the tuning of AI model parameters for the first example production activity data includes:
performing abnormal situation feature mining on the derived example production activity data according to the target AI analysis model, and determining derived abnormal situation features of the derived example production activity data;
performing linear feature conversion on a second forward abnormal situation feature or a second backward abnormal situation feature included in the derived abnormal situation features, and determining the second forward abnormal situation feature or the second backward abnormal situation feature after the linear feature conversion, wherein the second forward abnormal situation feature or the second backward abnormal situation feature after the linear feature conversion includes a plurality of abnormal situation feature members, and each abnormal situation feature member corresponds to one example production activity unit data in the derived example production activity data;
calculating an abnormal situation metric value for each abnormal situation feature member in the plurality of abnormal situation feature members, and obtaining third learning cost information based on the abnormal situation metric values of all the abnormal situation feature members, the feature parameter distribution of the derived abnormal situation features and the abnormal process distribution;
respectively calculating cross entropy learning cost information for each abnormal situation feature member in the plurality of abnormal situation feature members, and obtaining fourth learning cost information based on the cross entropy learning cost information of all the abnormal situation feature members, the feature parameter distribution of the derived abnormal situation features and the abnormal process distribution;
determining first target learning cost information of the target AI analysis model based on the third learning cost information and the fourth learning cost information;
performing abnormal situation feature mining on the first example production activity data according to the target AI analysis model, and determining first abnormal situation features of the first example production activity data;
determining second target learning cost information for the target AI analytical model based on the first abnormal situation characteristic and the example abnormal support probability value;
and adjusting and selecting parameter layer information of the target AI analysis model based on the first target learning cost information and the second target learning cost information.
For example, in one possible implementation of the first aspect, the first abnormal situation feature comprises a first forward abnormal situation feature and a first backward abnormal situation feature, and the example abnormal support probability value comprises a feature parameter distribution of an abnormal label of the abnormal production behavior, an abnormal process distribution, and an abnormal persistence of the abnormal production behavior in the first example production activity data;
the determining second target learning cost information of the target AI analysis model based on the first abnormal situation characteristic and the example abnormal support probability value includes:
determining first learning cost information based on the first forward abnormal situation feature, the abnormal continuation term of the abnormal production behavior, and the quantity of the first exemplar production activity data;
determining second learning cost information based on the first backward abnormal situation characteristics, the quantity of the first example production activity data, the characteristic parameter distribution and the abnormal process distribution of the abnormal label;
determining second target learning cost information of the target AI analysis model based on the first learning cost information and the second learning cost information.
For example, in a possible implementation manner of the first aspect, the first target learning cost information includes third learning cost information and fourth learning cost information, and the second target learning cost information includes first learning cost information and second learning cost information;
the adjusting and selecting parameter layer information of the target AI analysis model based on the first target learning cost information and the second target learning cost information comprises:
acquiring a first price influence value corresponding to the first learning cost information, a second price influence value corresponding to the second learning cost information, a third price influence value corresponding to the third learning cost information, and a fourth price influence value corresponding to the third learning cost information;
globally calculating the second target learning cost information and the first target learning cost information based on the first price influence value, the second price influence value, the third price influence value and the fourth price influence value, and determining target learning cost information;
and adjusting and selecting parameter layer information of the target AI analysis model based on the target learning cost information.
For example, in one possible implementation of the first aspect, the abnormal production behavior analysis model includes an abnormal production time-space domain feature mining branch and a feature aggregation branch;
the step of inputting the production activity data corresponding to each production node in the production node sequence into an abnormal production behavior analysis model for abnormal production behavior analysis to determine the probability representation of the abnormal production behavior of the production activity data includes:
according to the abnormal production time-space domain feature mining branch, performing abnormal production time-space domain feature mining on the production activity data, and determining initial abnormal production time-space domain features of the production activity data;
performing convolution feature extraction and sampling processing on the initial abnormal production time-space domain feature according to the feature aggregation branch, and determining a first abnormal production time-space domain feature;
compressing and exciting the initial abnormal production time-space domain features according to the feature aggregation branch, determining cost influence values corresponding to the initial abnormal production time-space domain features, cascading the initial abnormal production time-space domain features based on the cost influence values, and determining second abnormal production time-space domain features;
and aggregating the first abnormal production time-space domain feature and the second abnormal production time-space domain feature, and determining the abnormal production behavior probability representation of the production activity data.
In a second aspect, an embodiment of the present application provides a washing control system, including:
a processor;
a memory having stored therein a computer program that, when executed, implements the image processing-based metal workpiece detection and cleaning control method of the first aspect.
By adopting the technical scheme of any one aspect, the metal workpiece image data of the metal workpiece to be detected is obtained, the decision is made on the metal workpiece image data based on the preset parameter layer information adjustment and the selected metal workpiece detection and cleaning decision model, and the cleaning decision data of each target workpiece unit in the metal workpiece image data is determined, so that the metal workpiece to be detected is cleaned and controlled based on the cleaning decision data of each target workpiece unit in the metal workpiece image data, the cleaning decision data of each target workpiece unit in the metal workpiece image data is decided in an image data analysis mode, and then the metal workpiece to be detected is cleaned and controlled, and the reliability of the cleaning control process can be improved.
Drawings
Fig. 1 is a schematic flowchart illustrating steps of a method for controlling detection and cleaning of a metal workpiece based on image processing according to an embodiment of the present disclosure;
fig. 2 is a block diagram schematically illustrating a structure of a cleaning control system for executing the metal workpiece detection cleaning control method based on image processing in fig. 1 according to an embodiment of the present disclosure.
Detailed Description
It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Referring to fig. 1, an embodiment of the present application provides a method for controlling a metal workpiece inspection and cleaning based on image processing, which includes the following steps.
Step S110, metal workpiece image data for a metal workpiece to be detected is acquired.
Step S120, deciding the metal workpiece image data based on the preset parameter layer information optimization and the selected metal workpiece detection and cleaning decision model, and determining the cleaning decision data of each target workpiece unit in the metal workpiece image data, wherein the cleaning decision data comprises the cleaning time parameter and the cleaning mode parameter of each target workpiece unit. For example, the cleaning time parameter and the cleaning mode parameter may be determined according to the type and thickness of the organic coating of each target workpiece unit, that is, the organic coating thickness has a mapping relationship with the cleaning time parameter, and the organic coating type and the cleaning mode parameter have a mapping relationship.
And S130, cleaning and controlling the metal workpiece to be detected based on the cleaning decision data of each target workpiece unit in the metal workpiece image data.
For example, the metal workpiece to be detected may be subjected to cleaning control based on the cleaning time parameter and the cleaning manner parameter of each target workpiece unit.
Based on the above steps, in this embodiment, the metal workpiece image data for the metal workpiece to be detected is obtained, the decision is made on the metal workpiece image data based on the preset parameter layer information tuning and the selected metal workpiece detection and cleaning decision model, and the cleaning decision data of each target workpiece unit in the metal workpiece image data is determined, so that the metal workpiece to be detected is cleaned and controlled based on the cleaning decision data of each target workpiece unit in the metal workpiece image data, and the cleaning decision data of each target workpiece unit in the metal workpiece image data is determined in an image data analysis manner and then the metal workpiece to be detected is cleaned and controlled, so that the reliability of the cleaning control process can be improved.
In one possible implementation, for step S120, an example of model training of the metal workpiece detection and cleaning decision model may be referred to as the following exemplary embodiment.
A Process100 extracting a first sample workpiece image data sequence carrying prior knowledge and a second sample workpiece image data sequence not carrying prior knowledge;
the Process200 is used for deciding model learning cost information of an initialized metal workpiece detection and cleaning decision model corresponding to the initialized training cost decision model on the second example workpiece image data sequence not carrying the priori knowledge according to the initialized training cost decision model constructed by the first example workpiece image data sequence carrying the priori knowledge, and outputting the model training cost decision information;
for example, in some embodiments, the method includes obtaining sample workpiece image data to be decided, randomly selecting a part of the sample workpiece image data to be decided for a priori knowledge addition, and dividing the sample workpiece image data to be decided into a priori sample workpiece image data and a second sample workpiece image data, where the priori sample workpiece image data forms a first sample workpiece image data sequence carrying a priori knowledge, and the sample not carrying a priori knowledge forms a second sample workpiece image data sequence not carrying a priori knowledge; after obtaining the first example workpiece image data sequence carrying prior knowledge, an initialized metal workpiece detection cleaning decision model and an initialized training cost decision model may be constructed based on the first example workpiece image data sequence carrying prior knowledge; and performing training cost decision on the second example workpiece image data through the established initialized metal workpiece detection and cleaning decision model and the initialized training cost decision model to obtain model training cost decision information. For example, after the initialized metal workpiece inspection cleaning decision model and the initialized training cost decision model perform the training cost decision on the second example workpiece image data, the first cleaning control cost information of the second example workpiece image data and the model training cost decision information of the initialized metal workpiece inspection cleaning decision model are generated.
A Process300, extracting example workpiece image data from the second example workpiece image data sequence not carrying prior knowledge based on the model training cost decision information, and adding the prior knowledge to expand the first example workpiece image data sequence carrying prior knowledge;
for example, in some embodiments, the model-based training cost decision information is used to extract a portion of the sample workpiece image data from the second sample workpiece image data sequence that does not carry a priori knowledge for a priori knowledge addition, thereby expanding the first sample workpiece image data sequence that carries a priori knowledge. The method for expanding the first example workpiece image data sequence carrying the priori knowledge may be to extract a preset number of second example workpiece image data with the largest cleaning decision cost information in the model training cost decision information from the second example workpiece image data sequence not carrying the priori knowledge for artificial priori knowledge addition, or may also be to extract a preset number of second example workpiece image data with cleaning decision cost information greater than the target cleaning decision cost value in the model training cost decision information for prior knowledge addition. Therefore, the number of the priori carried example workpiece image data is increased after one round of priori knowledge addition every time, the initialized metal workpiece detection and cleaning decision model and the initialized training cost decision model can be trained based on the expanded priori carried example workpiece image data to update the initialized metal workpiece detection and cleaning decision model and the initialized training cost decision model, and therefore after each iteration period of training, the decision precision of the initialized metal workpiece detection and cleaning decision model is higher.
The Process400 performs iterative model parameter optimization on the initialized training cost decision model and the initialized metal workpiece detection and cleaning decision model according to the expanded first example workpiece image data sequence carrying the priori knowledge, and outputs the metal workpiece detection and cleaning decision model.
For example, in some embodiments, an initialization training cost decision model and an initialization metal workpiece detection and cleaning decision model are iteratively trained according to the expanded first example workpiece image data sequence carrying prior knowledge, the initialization metal workpiece detection and cleaning decision model is iteratively optimized, and the metal workpiece detection and cleaning decision model is output. After each iteration period, performing model evaluation once to judge whether the initialized metal workpiece detection and cleaning decision model meets the requirement of model deployment; and if the initialized metal workpiece detection and cleaning decision model meets the model deployment requirement, judging that the initialized metal workpiece detection and cleaning decision model is qualified, setting the initialized metal workpiece detection and cleaning decision model as a target metal workpiece detection and cleaning decision model, and directly applying the target metal workpiece detection and cleaning decision model to execute a metal workpiece detection and cleaning decision. If the initialized metal workpiece detection and cleaning decision model does not meet the model deployment requirement, updating model weight parameters in the initialized metal workpiece detection and cleaning decision model and the initialized training cost decision model, and then returning to the execution step: extracting a first example workpiece image data sequence carrying prior knowledge and a second example workpiece image data sequence not carrying prior knowledge, deciding model learning cost information of an initialized metal workpiece detection and cleaning decision model corresponding to the initialized training cost decision model on the second example workpiece image data sequence not carrying prior knowledge, outputting the model training cost decision information, continuing to expand the image data carrying prior example workpieces, and continuing to train the initialized metal workpiece detection and cleaning decision model based on the expanded image data carrying prior example workpieces until the initialized metal workpiece detection and cleaning decision model meets the requirement of model deployment. The model deployment requirement can be flexibly selected by a person skilled in the art based on actual design requirements, and the model deployment requirement is related to decision accuracy in a common situation.
For some possible design ideas, if the initialized metal workpiece detection and cleaning decision model meets the requirement of model deployment, the convergence precision of the model weight parameters of the initialized metal workpiece detection and cleaning decision model can be expected, and the deployment can be carried out; if the initialized metal workpiece detection and cleaning decision model does not meet the model deployment requirement, the convergence precision of the model weight parameters of the initialized metal workpiece detection and cleaning decision model can be considered to be not expected, and the initialized metal workpiece detection and cleaning decision model can not be deployed temporarily. In other embodiments, the model deployment requirements are related to the number of iterative training iterations.
Based on the steps, the initialized metal workpiece detection and cleaning decision model is trained, and the metal workpiece detection and cleaning decision is carried out on the metal workpiece to be detected based on the metal workpiece detection and cleaning decision model meeting the model deployment requirement, so that the accuracy of the cleaning decision result can be ensured, the manual priori knowledge is added with as little example image data as possible, the time cost of adding the manual priori knowledge and the processing cost in the optimization process of model parameter layer information are greatly reduced, and the training speed of the initialized metal workpiece detection and cleaning decision model is increased.
Further, the second example workpiece image data sequence without prior knowledge includes at least second example workpiece image data, the model training cost decision information includes at least cleaning decision cost information corresponding to the second example workpiece image data,
a1, the step of extracting example workpiece image data from the second example workpiece image data sequence not carrying prior knowledge and adding the prior knowledge based on the model training cost decision information to expand the first example workpiece image data sequence carrying prior knowledge specifically includes:
a2, extracting sample workpiece image data to be subjected to prior knowledge addition from the second sample workpiece image data sequence not carrying prior knowledge based on the size of each piece of cleaning decision cost information;
a3, performing priori knowledge addition on each sample workpiece image data to be subjected to the priori knowledge addition, and outputting each sample workpiece image data subjected to the priori knowledge addition;
a4, loading each sample workpiece image data added with the prior knowledge into the first sample workpiece image data sequence carrying the prior knowledge to expand the first sample workpiece image data sequence carrying the prior knowledge.
For example, in some embodiments, based on the size of each piece of cleaning decision cost information, each piece of example workpiece image data to be subjected to a priori knowledge addition is extracted from a second example workpiece image data sequence not carrying a priori knowledge; adding the prior knowledge to the sample workpiece image data to be added with the prior knowledge, and outputting the sample workpiece image data added with the prior knowledge; loading each of the a priori knowledge added exemplar workpiece image data into a first exemplar workpiece image data sequence carrying a priori knowledge to augment the first exemplar workpiece image data sequence carrying a priori knowledge. For example, the method for expanding the first example workpiece image data sequence carrying the priori knowledge may be to extract a preset number of second example workpiece image data with the largest cleaning decision cost information in the model training cost decision information from the second example workpiece image data sequence not carrying the priori knowledge for performing artificial priori knowledge addition, or may also be to extract a preset number of second example workpiece image data with cleaning decision cost information greater than the target cleaning decision cost value in the model training cost decision information for performing the priori knowledge addition.
For some possible design ideas, the initialized metal workpiece detection and cleaning decision Model is Model _ Target, the initialized training cost decision Model is Model _ Loss, and the initialized training cost decision Model is added to the initialized metal workpiece detection and cleaning decision Model.
For some possible design considerations, the step of calculating model training cost decision information of the second example workpiece image data by the initialized training cost decision model specifically includes:
b1, respectively converting the editing characteristic vectors of all the middle neural network units in the initialized metal workpiece detection and cleaning decision model into middle characteristic vectors of a training cost decision;
b2, fusing the intermediate feature vectors of the training cost decisions to obtain fused feature vectors;
and b3, passing the fusion characteristic vector through a full connection layer to obtain model training cost decision information.
For example, in some embodiments, the second example workpiece image data is used as an input to an initialized metal workpiece inspection and cleaning decision model, and an output of the initialized metal workpiece inspection and cleaning decision model is the first cleaning control cost information; the initialized metal workpiece detection and cleaning decision model can comprise an input unit, an output unit and a plurality of intermediate neural network units, and aiming at some possible design ideas, the intermediate neural network units comprise a data regularization layer, a convolution layer, a nonlinear excitation layer, a pooling layer and a connection layer. And respectively taking the edit characteristic vectors of each intermediate neural network unit of the initialized metal workpiece detection and cleaning decision model as the input of the initialized training cost decision model, and taking the output of the initialized training cost decision model as the model training cost decision information.
For some possible design considerations, the step of extracting the first example workpiece image data sequence carrying a priori knowledge and the second example workpiece image data sequence not carrying a priori knowledge is preceded by the method further comprising:
acquiring a candidate sample workpiece image data sequence to be decided and not carrying prior knowledge;
randomly extracting candidate sample workpiece image data without carrying a priori knowledge from the candidate sample workpiece image data sequence without carrying a priori knowledge for adding the priori knowledge, and dividing the candidate sample workpiece image data sequence without carrying the priori knowledge into a first sample workpiece image data sequence carrying the priori knowledge and a second sample workpiece image data sequence without carrying the priori knowledge;
and training and outputting an initialization training cost decision model and an initialization metal workpiece detection and cleaning decision model according to the first example workpiece image data sequence carrying the prior knowledge.
In order to improve the reliability of the training optimization, the second example workpiece image data with better training effect needs to be further screened. In some embodiments, for example, a candidate sample workpiece image data sequence to be decided, which does not carry a priori knowledge, is first obtained; then randomly extracting candidate example workpiece image data without carrying a priori knowledge from the candidate example workpiece image data sequence without carrying the a priori knowledge for adding the a priori knowledge, thereby dividing the candidate example workpiece image data sequence without carrying the a priori knowledge into a first example workpiece image data sequence with the a priori knowledge and a second example workpiece image data sequence without carrying the a priori knowledge; and training and outputting an initialization training cost decision model and an initialization metal workpiece detection and cleaning decision model according to a first example workpiece image data sequence carrying prior knowledge. The neural network model to be trained in the scheme comprises an initialized metal workpiece detection and cleaning decision model and an initialized training cost decision model. Therefore, besides the initialization metal workpiece detection and cleaning decision model for the cleaning decision of the metal workpiece to be detected is built and optimized, an initialization training cost decision model is added on the initialization metal workpiece detection and cleaning decision model, wherein the initialization training cost decision model is used for predicting the training cost of the initialization metal workpiece detection and cleaning decision model, and the deep network structure of deep learning occupies higher calculation cost, so that the initialization training cost decision model is designed to be much smaller than the initialization metal workpiece detection and cleaning decision model, and the initialization training cost decision model is attached to the initialization metal workpiece detection and cleaning decision model and is cooperated with the initialization metal workpiece detection and cleaning decision model for parameter optimization. Therefore, the calculation cost of the training cost decision for initializing the metal workpiece detection and cleaning decision model can be minimized. The initialization metal workpiece detection cleaning decision model and the initialization training cost decision model are convolution neural network models.
For some possible design ideas, the step of performing iterative model parameter optimization on the initialized training cost decision model and the initialized metal workpiece detection and cleaning decision model according to the expanded first example workpiece image data sequence carrying prior knowledge and outputting a metal workpiece detection and cleaning decision model specifically includes: calculating first cleaning control cost information of the initialized metal workpiece detection and cleaning decision model on the expanded first example workpiece image data sequence carrying prior knowledge; determining second cleaning control cost information of the initialized training cost decision model on the expanded first example workpiece image data sequence carrying prior knowledge based on the first cleaning control cost information; and performing iterative model parameter optimization on the initialized metal workpiece detection and cleaning decision model based on the first cleaning control cost information and the second cleaning control cost information, and outputting the metal workpiece detection and cleaning decision model.
For example, in some embodiments, first cleaning control cost information of an initialized metal workpiece inspection cleaning decision model on an augmented first example workpiece image data sequence carrying a priori knowledge is determined; determining second cleaning control cost information of the initialized training cost decision model on the expanded first example workpiece image data sequence carrying the priori knowledge according to the first cleaning control cost information; and outputting a metal workpiece detection and cleaning decision model based on the first cleaning control cost information and the second cleaning control cost information. For some possible design ideas, after iterative optimization is performed on the initialized metal workpiece detection and cleaning decision model each time, whether the initialized metal workpiece detection and cleaning decision model meets the model deployment requirement is judged, and if the initialized metal workpiece detection and cleaning decision model meets the model deployment requirement, the initialized metal workpiece detection and cleaning decision model is set as the metal workpiece detection and cleaning decision model.
For some possible design considerations, the step of determining, based on the first cleaning control cost information, second cleaning control cost information of the initialized training cost decision model on the augmented first example workpiece image data sequence with a priori knowledge specifically includes:
and according to the initialized training cost decision model, deciding model learning cost information of the initialized metal workpiece detection and cleaning decision model on the expanded first example workpiece image data sequence carrying the prior knowledge, and outputting second cleaning control cost information.
For example, in some embodiments, first cleaning control cost information of the a priori carried example workpiece image data may be calculated based on the target cleaning decision information of the a priori carried example workpiece image data and the a priori cleaning information added by the artificial a priori knowledge; the first cleaning control cost information is a real loss function value of the initialized metal workpiece detection cleaning decision model in the cleaning decision process. For the same prior image data carrying the example workpiece, the initial training cost decision model predicts a loss function value of the initial metal workpiece detection and cleaning decision model to obtain second cleaning control cost information.
For some possible design ideas, the step of performing iterative model parameter optimization on the initialized metal workpiece detection and cleaning decision model based on the first cleaning control cost information and the second cleaning control cost information, and outputting the metal workpiece detection and cleaning decision model specifically includes: determining target cleaning control cost information based on the first cleaning control cost information and the second cleaning control cost information; if the target cleaning control cost information meets the training termination requirement, taking the initialized metal workpiece detection cleaning decision model as the metal workpiece detection cleaning decision model; and if the target cleaning control cost information does not meet the training termination requirement, updating the model weight parameters of the initialized metal workpiece detection and cleaning decision model based on the model gradient calculated by the target cleaning control cost information, and returning to extract a first example workpiece image data sequence carrying prior knowledge and a second example workpiece image data sequence not carrying prior knowledge.
For some possible design ideas, the step of optimizing the initialized metal workpiece detection and cleaning decision model and outputting the initialized metal workpiece detection and cleaning decision model by performing iterative training on the initialized training cost decision model and the initialized metal workpiece detection and cleaning decision model which are constructed according to the priori carried example workpiece image data specifically includes: after each iterative training, judging whether the initialized metal workpiece detection and cleaning decision model meets the requirement of model deployment; if the initialized metal workpiece detection and cleaning decision model meets the requirement of model deployment, setting the initialized metal workpiece detection and cleaning decision model as the initialized metal workpiece detection and cleaning decision model; if the initialized metal workpiece detection and cleaning decision model does not meet the model deployment requirement, updating the initialized metal workpiece detection and cleaning decision model and the initialized training cost decision model, and returning to the execution step: a first sample workpiece image data sequence carrying a priori knowledge and a second sample workpiece image data sequence not carrying a priori knowledge are extracted.
For example, in some embodiments, an initialization training cost decision model and an initialization metal workpiece detection and cleaning decision model constructed according to a priori carried example workpiece image data are iteratively trained, and after each iteration cycle, model evaluation is performed to determine whether the initialization metal workpiece detection and cleaning decision model meets the requirement of model deployment; and if the initialized metal workpiece detection and cleaning decision model meets the requirement of model deployment, judging that the initialized metal workpiece detection and cleaning decision model can carry out model deployment, setting the initialized metal workpiece detection and cleaning decision model as the initialized metal workpiece detection and cleaning decision model, and directly applying the initialized metal workpiece detection and cleaning decision model to decide the cleaning decision information of the metal workpiece to be detected. If the initialized metal workpiece detection and cleaning decision model does not meet the model deployment requirement, returning to the execution step: extracting a first example workpiece image data sequence carrying prior knowledge and a second example workpiece image data sequence not carrying prior knowledge, performing training cost decision on the second example workpiece image data, outputting model training cost decision information, continuously updating the prior example workpiece image data, and continuously training an initialization metal workpiece detection and cleaning decision model based on the updated prior example workpiece image data until the initialization metal workpiece detection and cleaning decision model meets the requirement of model deployment.
For example, in a possible implementation, on the basis of the foregoing embodiment, the following steps may be further included in the application level.
And step S140, determining a cleaning control thermodynamic diagram of the corresponding metal workpiece production subarea according to a cleaning control data set formed by the cleaning control data of each metal workpiece to be detected. Wherein the cleaning control thermodynamic diagram is used for representing a distribution diagram formed by cleaning thermodynamic measurement values corresponding to each target workpiece unit in the corresponding metal workpiece production zone. The cleaning thermal metric may be determined by integrating a cleaning count and a cleaning time parameter (e.g., a product of the cleaning count and the cleaning time parameter) for each target workpiece unit.
Step S150, a production node sequence corresponding to a target workpiece unit meeting preset conditions in the cleaning control thermodynamic diagram is obtained, and whether abnormal production behaviors exist in production activity data corresponding to each production node in the production node sequence is analyzed, wherein the preset conditions include that a cleaning heating power metric value is larger than a preset heating power metric value.
And step S160, when the production activity data corresponding to any one production node has abnormal production behaviors, outputting the corresponding abnormal production node and the production node corresponding to the abnormal production node to the production service terminal for information prompt.
In step S150, the production activity data corresponding to each production node in the production node sequence may be input into an abnormal production behavior analysis model for abnormal production behavior analysis, and an abnormal production behavior probability representation of the production activity data is determined, where the abnormal production behavior analysis model is obtained by performing an abnormal characteristic AI training on the example production activity data of the metal workpiece.
For example, in one possible implementation, the model training process of the abnormal production behavior analysis model may include the following steps.
Step S101, obtaining an example production activity data sequence, wherein the example production activity data sequence comprises first example production activity data of the metal workpiece and derived example production activity data of the metal workpiece;
step S102, training a target AI analysis model based on performing an optimization update of AI model parameters on the derived example production activity data and performing an optimization update of AI model parameters on the first example production activity data, and when the target AI analysis model that reaches the parameter optimization output condition, taking the target AI analysis model that reaches the parameter optimization output condition as an abnormal production behavior analysis model, where the abnormal production behavior analysis model is used to analyze an abnormal support probability value of an abnormal production behavior in the production activity data of the metal workpiece.
For instance, the example production activity data sequence further includes example abnormal support probability values for abnormal production behavior in the first example production activity data; training a targeted AI analysis model based on the tuning updates of AI model parameters to the derived example production activity data and the tuning updates of AI model parameters to the first example production activity data, comprising: performing abnormal situation feature mining on the derived example production activity data according to the target AI analysis model, and determining derived abnormal situation features of the derived example production activity data; performing linear feature conversion on a second forward abnormal situation feature or a second backward abnormal situation feature included in the derived abnormal situation features, and determining the second forward abnormal situation feature or the second backward abnormal situation feature after the linear feature conversion, wherein the second forward abnormal situation feature or the second backward abnormal situation feature after the linear feature conversion includes a plurality of abnormal situation feature members, and each abnormal situation feature member corresponds to one example production activity unit data in the derived example production activity data; calculating an abnormal situation characteristic metric value for each abnormal situation characteristic member in the plurality of abnormal situation characteristic members respectively, and obtaining third learning cost information based on the abnormal situation metric values of all the abnormal situation characteristic members, the characteristic parameter distribution of the derived abnormal situation characteristics and the abnormal process distribution; respectively calculating cross entropy learning cost information for each abnormal situation feature member in the abnormal situation feature members, and obtaining fourth learning cost information based on the cross entropy learning cost information of all the abnormal situation feature members, the feature parameter distribution of the derived abnormal situation features and the abnormal process distribution; determining first target learning cost information of the target AI analysis model based on the third learning cost information and the fourth learning cost information; performing abnormal situation feature mining on the first example production activity data according to the target AI analysis model, and determining first abnormal situation features of the first example production activity data; determining second target learning cost information for the target AI analytical model based on the first abnormal situation characteristic and the example abnormal support probability value; and adjusting and selecting parameter layer information of the target AI analysis model based on the first target learning cost information and the second target learning cost information.
For example, in one possible implementation, the first abnormal situation feature comprises a first forward abnormal situation feature and a first backward abnormal situation feature, and the example abnormal support probability value comprises a feature parameter distribution of an abnormal label of the abnormal production behavior, an abnormal process distribution, and an abnormal duration of the abnormal production behavior in the first example production activity data; determining second target learning cost information for the target AI analytical model based on the first anomaly situational features and the example anomaly support probability values, including: determining first learning cost information based on the first forward abnormal situation feature, the abnormal continuation of the abnormal production behavior, and the quantity of the first exemplar production activity data; determining second learning cost information based on the first backward abnormal situation characteristics, the quantity of the first example production activity data, the characteristic parameter distribution and the abnormal process distribution of the abnormal label; determining second target learning cost information of the target AI analysis model based on the first learning cost information and the second learning cost information.
For example, in one possible implementation, the first target learning cost information includes third learning cost information and fourth learning cost information, and the second target learning cost information includes first learning cost information and second learning cost information; adjusting and selecting parameter layer information of the target AI analysis model based on the first target learning cost information and the second target learning cost information, comprising: acquiring a first price influence value corresponding to the first learning cost information, a second price influence value corresponding to the second learning cost information, a third price influence value corresponding to the third learning cost information, and a fourth price influence value corresponding to the third learning cost information;
globally calculating the second target learning cost information and the first target learning cost information based on the first price influence value, the second price influence value, the third price influence value and the fourth price influence value, and determining target learning cost information; and adjusting and selecting parameter layer information of the target AI analysis model based on the target learning cost information.
For example, in one possible implementation, the abnormal production behavior analysis model includes an abnormal production time-space domain feature mining branch and a feature aggregation branch; inputting the production activity data corresponding to each production node in the production node sequence into an abnormal production behavior analysis model for abnormal production behavior analysis, and determining the probability representation of the abnormal production behavior of the production activity data, wherein the method comprises the following steps: according to the abnormal production time-space domain feature mining branch, performing abnormal production time-space domain feature mining on the production activity data, and determining initial abnormal production time-space domain features of the production activity data; performing convolution feature extraction and sampling processing on the initial abnormal production time-space domain feature according to the feature aggregation branch to determine a first abnormal production time-space domain feature; compressing and exciting the initial abnormal production time-space domain features according to the feature aggregation branch, determining cost influence values corresponding to the initial abnormal production time-space domain features, cascading the initial abnormal production time-space domain features based on the cost influence values, and determining second abnormal production time-space domain features; and aggregating the first abnormal production time-space domain characteristics and the second abnormal production time-space domain characteristics, and determining the probability representation of the abnormal production behaviors of the production activity data.
Further, as shown in fig. 2, the present embodiment also provides a cleaning control system, and the cleaning control system 100 may have a relatively large difference due to different configurations or performances, and may include one or more Central Processing Units (CPUs) 112 (e.g., one or more processors) and a memory 111. Wherein the memory 111 may be a transient storage or a persistent storage. The program stored in the memory 111 may include one or more modules, each of which may include a sequence of instructions operating on the washing control system 100. Further, the central processor 112 may be configured to communicate with the memory 111 to execute a series of command operations in the memory 111 on the washing control system 100.
The washing control system 100 may also include one or more power supplies, one or more communication units 113, one or more pass-to-output interfaces, and/or one or more operating systems, such as Windows Server, mac OS XTM, unixTM, linuxTM, freeBSDTM, and the like.
In addition, a storage medium is provided in an embodiment of the present application, and the storage medium is used for storing a computer program, and the computer program is used for executing the method provided in the embodiment.
The embodiment of the present application also provides a computer program product including instructions, which when run on a computer, causes the computer to execute the method provided by the above embodiment.
Those of ordinary skill in the art will understand that: all or part of the steps of implementing the method embodiments may be implemented by hardware associated with program instructions, where the program may be stored in a computer-readable storage medium, and when executed, performs the steps including the method embodiments; and the aforementioned storage medium may be at least one of the following media: various media that can store program codes, such as a Read-only Memory (ROM), a RAM, a magnetic disk, or an optical disk.
It should be noted that, in the present specification, each embodiment is described in a progressive manner, and the same and similar parts among the embodiments may be referred to each other, and each embodiment is intended to be different from other embodiments. In particular, the apparatus and system embodiments, because they are substantially similar to the method embodiments, are described in a relatively simple manner, and reference may be made to some of the descriptions of the method embodiments for related points. The above-described embodiments of the apparatus and system are merely illustrative, and the units described as separate parts may or may not be physically separate, and the parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.
The above description is only one specific embodiment of the present application, but the scope of the present application is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present application should be included in the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims (10)

1. A metal workpiece detection and cleaning control method based on image processing is applied to a cleaning control system and is characterized by comprising the following steps:
acquiring metal workpiece image data aiming at a metal workpiece to be detected;
performing decision making on the metal workpiece image data based on parameter layer information tuning in advance and a selected metal workpiece detection and cleaning decision model, and determining cleaning decision data of each target workpiece unit in the metal workpiece image data, wherein the cleaning decision data comprises a cleaning time parameter and a cleaning mode parameter of each target workpiece unit;
cleaning control is carried out on the metal workpiece to be detected based on cleaning decision data of each target workpiece unit in the metal workpiece image data;
the method further comprises the following steps: extracting a first sample workpiece image data sequence carrying prior knowledge and a second sample workpiece image data sequence not carrying prior knowledge;
after training cost decision is performed on the second example workpiece image data based on the initialized metal workpiece detection and cleaning decision model and the initialized training cost decision model, first cleaning control cost information of the second example workpiece image data and model training cost decision information of the initialized metal workpiece detection and cleaning decision model are generated; the initialization training cost decision model is attached to the initialization metal workpiece detection and cleaning decision model, is optimized in cooperation with the initialization metal workpiece detection and cleaning decision model parameters, and according to the initialization training cost decision model constructed by the first example workpiece image data sequence carrying prior knowledge, decides model learning cost information of the initialization metal workpiece detection and cleaning decision model corresponding to the initialization training cost decision model on the second example workpiece image data sequence not carrying prior knowledge, and outputs model training cost decision information;
based on the model training cost decision information, extracting example workpiece image data from the second example workpiece image data sequence which does not carry a priori knowledge for adding the priori knowledge so as to expand the first example workpiece image data sequence carrying the priori knowledge;
and performing iterative model parameter optimization on the initialized training cost decision model and the initialized metal workpiece detection and cleaning decision model according to the expanded first example workpiece image data sequence carrying prior knowledge, and outputting the metal workpiece detection and cleaning decision model.
2. The method according to claim 1, wherein the second example workpiece image data sequence not carrying prior knowledge at least comprises second example workpiece image data, the model training cost decision information at least comprises cleaning decision cost information corresponding to the second example workpiece image data, and the step of extracting example workpiece image data from the second example workpiece image data sequence not carrying prior knowledge for prior knowledge addition based on the model training cost decision information to expand the first example workpiece image data sequence carrying prior knowledge comprises:
extracting example workpiece image data to be subjected to prior knowledge addition from the second example workpiece image data sequence which does not carry prior knowledge based on the size of each piece of cleaning decision cost information;
adding the prior knowledge to the sample workpiece image data to be added with the prior knowledge, and outputting the sample workpiece image data added with the prior knowledge;
loading each of the a priori knowledge added sample workpiece image data into the a priori knowledge carried first sample workpiece image data sequence to augment the a priori knowledge carried first sample workpiece image data sequence.
3. The method according to claim 2, wherein the step of determining model learning cost information of the initialized metal workpiece detection and cleaning decision model corresponding to the initialized training cost decision model on the second example workpiece image data sequence not carrying the priori knowledge and outputting the model training cost decision information includes:
respectively converting the editing characteristic vectors of all the middle neural network units in the initialized metal workpiece detection and cleaning decision model into middle characteristic vectors of a training cost decision;
fusing the intermediate feature vectors of the training cost decisions to obtain fused feature vectors;
and carrying out full connection on the fusion characteristic vectors and outputting the model training cost decision information.
4. The image processing-based metal workpiece inspection cleaning control method of claim 1, wherein prior to the step of extracting a first sample workpiece image data sequence carrying a priori knowledge and a second sample workpiece image data sequence not carrying a priori knowledge, the method further comprises:
acquiring a candidate sample workpiece image data sequence to be decided, which does not carry prior knowledge;
randomly extracting candidate sample workpiece image data without carrying a priori knowledge from the candidate sample workpiece image data sequence without carrying a priori knowledge for adding the priori knowledge, and dividing the candidate sample workpiece image data sequence without carrying the priori knowledge into a first sample workpiece image data sequence carrying the priori knowledge and a second sample workpiece image data sequence without carrying the priori knowledge;
and training and outputting an initialization training cost decision model and an initialization metal workpiece detection and cleaning decision model according to the first example workpiece image data sequence carrying the prior knowledge.
5. The method according to claim 1, wherein the step of performing iterative model parameter optimization on the initialized training cost decision model and the initialized metal workpiece detection and cleaning decision model according to the expanded first example workpiece image data sequence with prior knowledge and outputting a metal workpiece detection and cleaning decision model specifically comprises:
calculating first cleaning control cost information of the initialized metal workpiece detection and cleaning decision model on the expanded first example workpiece image data sequence carrying prior knowledge;
determining second cleaning control cost information of the initialized training cost decision model on the expanded first example workpiece image data sequence carrying prior knowledge based on the first cleaning control cost information;
and performing iterative model parameter optimization on the initialized metal workpiece detection and cleaning decision model based on the first cleaning control cost information and the second cleaning control cost information, and outputting the metal workpiece detection and cleaning decision model.
6. The method of claim 5, wherein the step of determining second cleaning control cost information of the initialized trained cost decision model on the augmented first example workpiece image data sequence with a priori knowledge based on the first cleaning control cost information comprises:
and according to the initialized training cost decision model, deciding model learning cost information of the initialized metal workpiece detection and cleaning decision model on the expanded first example workpiece image data sequence carrying prior knowledge, and outputting second cleaning control cost information.
7. The image-processing-based metal workpiece detection and cleaning control method according to claim 5, wherein the step of performing iterative model parameter optimization on the initialized metal workpiece detection and cleaning decision model based on the first cleaning control cost information and the second cleaning control cost information and outputting the metal workpiece detection and cleaning decision model specifically comprises:
determining target cleaning control cost information based on the first cleaning control cost information and the second cleaning control cost information;
if the target cleaning control cost information meets the training termination requirement, taking the initialized metal workpiece detection cleaning decision model as the metal workpiece detection cleaning decision model;
and if the target cleaning control cost information does not meet the training termination requirement, updating the model weight parameters of the initialized metal workpiece detection and cleaning decision model based on the model gradient calculated by the target cleaning control cost information, and returning to extract a first example workpiece image data sequence carrying prior knowledge and a second example workpiece image data sequence not carrying prior knowledge.
8. The image processing-based metal workpiece detection and cleaning control method according to any one of claims 1-7, characterized by further comprising:
determining a cleaning control thermodynamic diagram of a corresponding metal workpiece production partition according to a cleaning control data set formed by cleaning control data of each metal workpiece to be detected, wherein the cleaning control thermodynamic diagram is used for representing a distribution diagram formed by cleaning thermodynamic measurement values corresponding to each target workpiece unit in the corresponding metal workpiece production partition;
acquiring a production node sequence corresponding to a target workpiece unit meeting a preset condition in the cleaning control thermodynamic diagram, and analyzing whether production activity data corresponding to each production node in the production node sequence has abnormal production behavior, wherein the preset condition comprises that a cleaning thermodynamic metric value is larger than a preset thermodynamic metric value;
when abnormal production behaviors exist in the production activity data corresponding to any one production node, outputting the corresponding abnormal production node and the production node corresponding to the abnormal production node to a production service terminal for information prompt;
wherein, the step of analyzing whether the production activity data corresponding to each production node in the production node sequence has abnormal production behavior includes:
inputting the production activity data corresponding to each production node in the production node sequence into an abnormal production behavior analysis model for abnormal production behavior analysis, and determining the probability expression of the abnormal production behavior of the production activity data, wherein the abnormal production behavior analysis model is obtained by performing abnormal characteristic AI training on the sample production activity data of the metal workpiece;
before the step of inputting the production activity data corresponding to each production node in the production node sequence into an abnormal production behavior analysis model for abnormal production behavior analysis and determining the probability representation of the abnormal production behavior of the production activity data, the method further includes:
obtaining an example production campaign data sequence comprising first example production campaign data for the metal workpiece and derived example production campaign data for the metal workpiece;
training a target AI analysis model based on carrying out optimization updating on AI model parameters on the derived example production activity data and carrying out optimization updating on AI model parameters on the first example production activity data, and when the target AI analysis model reaching the parameter optimization output condition reaches the parameter optimization output condition, taking the target AI analysis model reaching the parameter optimization output condition as an abnormal production behavior analysis model used for analyzing an abnormal support probability value of abnormal production behaviors in the production activity data of the metal workpiece;
the example production activity data sequence further includes example abnormal support probability values for abnormal production behavior in the first example production activity data;
the training of the subject AI analytical model based on the tuning updates of AI model parameters to the derived example production activity data and the tuning updates of AI model parameters to the first example production activity data includes:
performing abnormal situation feature mining on the derived example production activity data according to the target AI analysis model, and determining derived abnormal situation features of the derived example production activity data;
performing linear feature conversion on a second forward abnormal situation feature or a second backward abnormal situation feature included in the derived abnormal situation features, and determining the second forward abnormal situation feature or the second backward abnormal situation feature after the linear feature conversion, wherein the second forward abnormal situation feature or the second backward abnormal situation feature after the linear feature conversion includes a plurality of abnormal situation feature members, and each abnormal situation feature member corresponds to one example production activity unit data in the derived example production activity data;
calculating an abnormal situation characteristic metric value for each abnormal situation characteristic member in the plurality of abnormal situation characteristic members respectively, and obtaining third learning cost information based on the abnormal situation metric values of all the abnormal situation characteristic members, the characteristic parameter distribution of the derived abnormal situation characteristics and the abnormal process distribution;
respectively calculating cross entropy learning cost information for each abnormal situation feature member in the abnormal situation feature members, and obtaining fourth learning cost information based on the cross entropy learning cost information of all the abnormal situation feature members, the feature parameter distribution of the derived abnormal situation features and the abnormal process distribution;
determining first target learning cost information of the target AI analysis model based on the third learning cost information and the fourth learning cost information;
performing abnormal situation feature mining on the first example production activity data according to the target AI analysis model, and determining first abnormal situation features of the first example production activity data;
determining second target learning cost information for the target AI analytical model based on the first abnormal situation characteristic and the example abnormal support probability value;
adjusting and selecting parameter layer information of the target AI analysis model based on the first target learning cost information and the second target learning cost information;
the first abnormal situation feature comprises a first forward abnormal situation feature and a first backward abnormal situation feature, and the example abnormal support probability value comprises a feature parameter distribution of an abnormal label of the abnormal production behavior, an abnormal process distribution and an abnormal continuation term of the abnormal production behavior in the first example production activity data.
9. The image-processing-based metal workpiece inspection and cleaning control method of claim 8, wherein the determining a second target learning cost information of the target AI analytical model based on the first anomaly situational characteristics and the example anomaly support probability values comprises:
determining first learning cost information based on the first forward abnormal situation feature, the abnormal continuation term of the abnormal production behavior, and the quantity of the first exemplar production activity data;
determining second learning cost information based on the first backward abnormal situation characteristics, the quantity of the first example production activity data, the characteristic parameter distribution and the abnormal process distribution of the abnormal label;
determining second target learning cost information of the target AI analysis model based on the first learning cost information and the second learning cost information;
the first target learning cost information comprises third learning cost information and fourth learning cost information, and the second target learning cost information comprises first learning cost information and second learning cost information;
the adjusting and selecting parameter layer information of the target AI analysis model based on the first target learning cost information and the second target learning cost information comprises:
acquiring a first price influence value corresponding to the first learning cost information, a second price influence value corresponding to the second learning cost information, a third price influence value corresponding to the third learning cost information, and a fourth price influence value corresponding to the third learning cost information;
globally calculating the second target learning cost information and the first target learning cost information based on the first price influence value, the second price influence value, the third price influence value and the fourth price influence value, and determining target learning cost information;
adjusting and selecting parameter layer information of the target AI analysis model based on the target learning cost information;
the abnormal production behavior analysis model comprises an abnormal production time-space domain feature mining branch and a feature aggregation branch;
the step of inputting the production activity data corresponding to each production node in the production node sequence into an abnormal production behavior analysis model for abnormal production behavior analysis to determine the probability representation of the abnormal production behavior of the production activity data includes:
according to the abnormal production time-space domain feature mining branch, performing abnormal production time-space domain feature mining on the production activity data, and determining initial abnormal production time-space domain features of the production activity data;
performing convolution feature extraction and sampling processing on the initial abnormal production time-space domain feature according to the feature aggregation branch to determine a first abnormal production time-space domain feature;
compressing and exciting the initial abnormal production time-space domain features according to the feature aggregation branch, determining cost influence values corresponding to the initial abnormal production time-space domain features, cascading the initial abnormal production time-space domain features based on the cost influence values, and determining second abnormal production time-space domain features;
and aggregating the first abnormal production time-space domain feature and the second abnormal production time-space domain feature, and determining the abnormal production behavior probability representation of the production activity data.
10. A purge control system, comprising:
a processor;
a memory having stored therein a computer program that, when executed, implements the image processing-based metal workpiece inspection and cleaning control method of any one of claims 1-9.
CN202211350654.8A 2022-10-31 2022-10-31 Metal workpiece detection and cleaning control method and system based on image processing Active CN115409836B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211350654.8A CN115409836B (en) 2022-10-31 2022-10-31 Metal workpiece detection and cleaning control method and system based on image processing

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211350654.8A CN115409836B (en) 2022-10-31 2022-10-31 Metal workpiece detection and cleaning control method and system based on image processing

Publications (2)

Publication Number Publication Date
CN115409836A CN115409836A (en) 2022-11-29
CN115409836B true CN115409836B (en) 2023-01-10

Family

ID=84168162

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211350654.8A Active CN115409836B (en) 2022-10-31 2022-10-31 Metal workpiece detection and cleaning control method and system based on image processing

Country Status (1)

Country Link
CN (1) CN115409836B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116679639B (en) * 2023-05-26 2024-01-05 广州市博煌节能科技有限公司 Optimization method and system of metal product production control system
CN117169449B (en) * 2023-10-31 2023-12-29 江苏权正检验检测有限公司 Food decomposition and cleaning method and system for food detection

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105127150A (en) * 2015-08-14 2015-12-09 四川大学 Laser cleaning system based on robot control and cleaning method thereof
CN107340302A (en) * 2017-07-06 2017-11-10 武汉翔明激光科技有限公司 A kind of cleaning quality monitoring device and method based on laser cleaner
CN108906763A (en) * 2018-05-16 2018-11-30 深圳仕上电子科技有限公司 Utilize the method for acetone detergent solution removing workpiece surface film
CN110836896A (en) * 2019-10-14 2020-02-25 深圳信息职业技术学院 Laser cleaning detection equipment and laser cleaning detection method
CN111539552A (en) * 2020-03-19 2020-08-14 卡斯柯信号有限公司 Method for predicting life cycle of electronic component based on semiconductor market data
JP6887033B1 (en) * 2020-02-12 2021-06-16 Dmg森精機株式会社 Image processing equipment, machine tools and image processing methods

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2019161606A (en) * 2018-03-16 2019-09-19 オリンパス株式会社 Mobile imaging system, learning method, mobile imaging device, information acquisition control device, information acquisition control method, and information acquisition control program
US20220300791A1 (en) * 2021-03-19 2022-09-22 Micron Technology, Inc. Systems and methods for training machine learning models for denoising images

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105127150A (en) * 2015-08-14 2015-12-09 四川大学 Laser cleaning system based on robot control and cleaning method thereof
CN107340302A (en) * 2017-07-06 2017-11-10 武汉翔明激光科技有限公司 A kind of cleaning quality monitoring device and method based on laser cleaner
CN108906763A (en) * 2018-05-16 2018-11-30 深圳仕上电子科技有限公司 Utilize the method for acetone detergent solution removing workpiece surface film
CN110836896A (en) * 2019-10-14 2020-02-25 深圳信息职业技术学院 Laser cleaning detection equipment and laser cleaning detection method
JP6887033B1 (en) * 2020-02-12 2021-06-16 Dmg森精機株式会社 Image processing equipment, machine tools and image processing methods
CN111539552A (en) * 2020-03-19 2020-08-14 卡斯柯信号有限公司 Method for predicting life cycle of electronic component based on semiconductor market data

Also Published As

Publication number Publication date
CN115409836A (en) 2022-11-29

Similar Documents

Publication Publication Date Title
CN115409836B (en) Metal workpiece detection and cleaning control method and system based on image processing
CN108229667B (en) Trimming based on artificial neural network classification
EP3407267A1 (en) Deep learning network architecture optimization for uncertainty estimation in regression
KR102457974B1 (en) Method and apparatus for searching new material
KR20190048895A (en) Method For optimizing hyper-parameter automatically and Apparatus thereof
CN110096938B (en) Method and device for processing action behaviors in video
CN113037577B (en) Network traffic prediction method, device and computer readable storage medium
Imani et al. Multiple model adaptive controller for partially-observed Boolean dynamical systems
CN113780466A (en) Model iterative optimization method and device, electronic equipment and readable storage medium
JP2018528511A (en) Optimizing output efficiency in production systems
US20200167660A1 (en) Automated heuristic deep learning-based modelling
CN110866602A (en) Method and device for integrating multitask model
KR102464891B1 (en) Apparatus and method for controlling metal part intelligent processing
CN112884569A (en) Credit assessment model training method, device and equipment
US11003823B2 (en) Re-design of analog circuits
CN113222843B (en) Image restoration method and related equipment thereof
CN114650552B (en) Method and apparatus for anomaly detection in a network
WO2011085819A1 (en) A machine-learning system and a method for determining different operating points in such a system
KR101151013B1 (en) Method for evaluating performance of tire
JP7424373B2 (en) Analytical equipment, analytical methods and analytical programs
CN113869033A (en) Graph neural network sentence sequencing method integrated with iterative sentence pair relation prediction
EP2909687B1 (en) System testing algorithm and apparatus
JP2022138758A (en) Information processing apparatus, information processing method, and program
CN115145904B (en) Big data cleaning method and big data acquisition system for AI cloud computing training
CN112052626B (en) Automatic design system and method for neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CP01 Change in the name or title of a patent holder
CP01 Change in the name or title of a patent holder

Address after: 518103 Floor 2 and 4 of Building B, Huixin Industrial Park, Xinhe Community, Fuhai Street, Bao'an District, Shenzhen, Guangdong Province, and Floor 1 of Building C

Patentee after: Shenzhen Shishang Electronic Technology Co.,Ltd.

Address before: 518103 Floor 2 and 4 of Building B, Huixin Industrial Park, Xinhe Community, Fuhai Street, Bao'an District, Shenzhen, Guangdong Province, and Floor 1 of Building C

Patentee before: SHENZHEN SHISHANG ELECTRONIC TECHNOLOGY CO.,LTD.