CN120832106B

CN120832106B - Edge calculation storage device for self-learning of printing process parameters

Info

Publication number: CN120832106B
Application number: CN202511247788.0A
Authority: CN
Inventors: 庄云婷; 何骏豪; 李明杰
Original assignee: Quanzhou Daewoo 3d Printing Technology Co ltd
Current assignee: Quanzhou Daewoo 3d Printing Technology Co ltd
Priority date: 2025-09-03
Filing date: 2025-09-03
Publication date: 2026-01-20
Anticipated expiration: 2045-09-03
Also published as: CN120832106A

Abstract

The invention belongs to the technical field of printing equipment optimization and edge calculation, and particularly relates to an edge calculation storage device for printing process parameter self-learning. Firstly, the invention provides a system based on edge calculation, which combines data acquired by a sensor to realize efficient data management and storage through digital processing and encryption storage, secondly, dynamic optimization of a printing process and real-time adaptation of model update are ensured by utilizing local model training and federal learning technology, and finally, intelligent management and efficient operation and maintenance of printing equipment are realized through global weight update and intelligent adjustment of a local model by adopting a safe interaction module. By the technical scheme, the stability and quality consistency of the printing process are obviously improved, manual intervention is reduced, and the response speed and the data management capability of the system are improved.

Description

Edge calculation storage device for self-learning of printing process parameters

Technical Field

The invention relates to the technical field of printing equipment optimization and edge calculation, in particular to an edge calculation storage device for printing process parameter self-learning.

Background

In the operation process of the printing equipment, a large amount of process-related data including various physical quantities such as temperature, humidity, pressure and speed can be generated, the data not only reflects the working state of the printing equipment, but also directly influences the precision and quality of a finished product, in order to realize effective management of the process, the data are required to be stored, processed and analyzed, and stored together with information such as model parameters and the like, so that support is provided for subsequent process adjustment, model training and quality tracing, common storage media comprise a solid state disk and a memory card, efficient data recording and management can be realized at the equipment end, the information integrity and safety in the operation process are ensured, and a solid foundation is provided for optimizing and intelligent development of the printing process through systematic storage and utilization of the process data.

However, the data storage and processing in the prior art mainly depend on a centralized server, so that the data transmission delay is higher, especially in industrial printing scenes with strict real-time requirements, the demands of low delay and high efficiency cannot be met, the centralized processing makes equipment far away from the server have strong network dependence and are easily influenced by network problems or communication interruption, secondly, the existing printing equipment usually depends on manual adjustment or preset fixed parameters, so that the process adjustment in the printing process is rough, the material change, the environmental fluctuation or the equipment aging cannot be automatically adapted, the production efficiency is low, the product consistency is poor, and finally, the model update in the prior art depends on manual intervention or frequent manual calibration, so that the operation complexity is increased, the model is outdated or does not meet the actual running condition, and flexible and real-time adaptation and optimization cannot be realized.

Therefore, the invention provides edge computing storage equipment for self-learning of printing process parameters.

Disclosure of Invention

In order to solve the problems of low data processing and storage efficiency, poor self-adaptive capability of process parameters and complex model updating and adaptation of the prior printing and storing device technology in the prior background art, the invention aims to provide the edge computing and storing device for self-learning of the printing process parameters.

In order to achieve the above purpose, the present invention adopts the following technical scheme:

a printing process parameter self-learning edge calculation storage device comprising:

The system comprises a local acquisition and storage module, a Central Processing Unit (CPU), a data acquisition board and a flash memory chip, wherein the local acquisition and storage module consists of a sensor, the data acquisition board and the flash memory chip, and output signals of the sensor are integrated into parameter data files by the CPU after being digitally processed by the data acquisition board and are stored in the flash memory chip in an encrypted manner;

The hierarchical data management module is used for cooperatively working by the central processing unit, the flash memory chip and the EEPROM, the central processing unit carries out importance level division on the parameter data files according to a set period, and after threshold comparison, the parameter data files are respectively stored in the flash memory chip or the EEPROM;

M3, a local model training module acquires the parameter data file, invokes the central processing unit to perform model training on the parameter data file, generates a model parameter abstract, and synchronously writes the model parameter abstract into the EEPROM;

And M4, an interaction module is used for uploading the model parameter abstract in the EEPROM to the central processing unit in a set time interval, and simultaneously receiving the global model weight issued by the central processing unit, and the central processing unit is used for covering the local model weight with the global model weight after searching and matching.

Further, the sensor comprises a temperature sensor, a humidity sensor, a pressure sensor and a speed sensor, wherein the temperature sensor is used for monitoring the surface temperature and the environment temperature of the printing head and the substrate, the humidity sensor is used for monitoring the humidity of the environment and the substrate, the pressure sensor is used for monitoring the pressure of the printing head, and the speed sensor is used for monitoring the printing speed of the printer;

First, the sensor converts the acquired physical quantity into an analog electrical signal Outputting, then, the data acquisition board sets a fixed sampling periodAnd periodically sampling from the analog signal during the period, the sampling process is as follows:

At the position of Sampling at the moment and recording the sampled analog signal asRepresenting the analog electrical signal inSignal value of time of day, andWherein, the method comprises the steps of,;Is the initial time; for sampling index, representing different sampling time;

the analog signal after sampling Conversion to digital signals by ADC;

Based on the analog signalAnd the digital signalCalculating quantization errorThe formula is, wherein,The maximum voltage value allowed for the ADC input,Inputting a permissible minimum voltage value for the ADC;

Setting the maximum quantization error as And (2) andIf (1)The original ADC bit number is maintainedUnchanged, ifThen the ADC bit number needs to be increasedTo reduce the error, the calculation formula is:

Wherein, the Is the minimum number of bits of the ADC;

All of the digital signals Are integrated into a parameter data file according to time sequenceThe formula is:

。

further, based on the parameter data file Calculating the variation amplitudeThe formula is:

And then based on the amplitude of the change Calculated to obtain the firstImportance level of each of the parameter data filesThe formula is:

Wherein, the Data sources representing different types of sensor data;And Are weight coefficients, and sequentially correspond to the change amplitude and the influence degree of the data source on the importance level;

Ranking the importance levels And set a threshold valueComparing, and determining storage positions for dividing the parameter data files as follows:

Wherein, the storage position corresponding to 1 is a flash memory chip, and the storage position corresponding to 2 is an EEPROM.

Further, based on the parameter data in the parameter data file, calculating a mean value of each featureAnd standard deviationComposition vector:

Wherein, the Representing dimensions asEach element belonging to a real set;

And the central processing unit performs model training on the parameter data in the standardized parameter data file through ridge regression, wherein the training process is expressed as follows:

Wherein, the Representing a minimized loss function, representing a model training target; Is regression coefficient vector with the size of Representing the contribution of each input feature to the target output; is a bias term of the size The method is used for adjusting the reference value output by the model; the regularization parameters are used for controlling the complexity of regression coefficients and preventing overfitting; For the target output vector, the size is Representing the actual output of the printing process; the square representing the euclidean norm, the sum of squares representing the vector, for measuring the error; is a standardized feature matrix; Is an offset term For expanding to a target output vectorVectors of consistent dimensions;

Based on the minimized loss function Calculating regression coefficientsIs calculated by the following formula:

Wherein, the Is the estimated value of regression coefficient with the size of;Is a unit matrix with the size of;Is an input feature matrix; For inputting feature matrix Is a transposed matrix of (a);

Based on the minimized loss function Calculating bias termsIs calculated by the following formula:

Wherein, the An estimated value for the bias term; The average value of the target output is ;

After training is completed, the training residual is calculated by the following formula:

Wherein, the Is a residual vector of the size ofRepresenting the difference between the predicted value and the actual value; Representing the bias term Extended to and predicted output vectorColumn vectors of consistent dimensions;

the performance of the ridge regression model was then evaluated using the following criteria:

Wherein, the The mean square error is used for measuring the difference between the model predicted value and the true value; for determining coefficients, the method is used for measuring model fitting goodness;

Generating the model parameter abstract Expressed as:

Finally abstracting the model parameters Synchronously writing into EEPROM for storage.

Further, the edge device sets the time intervalUploading the model parameter abstract in the EEPROM to a central processing unit, and calculating the time interval formula asWherein, the method comprises the steps of,The starting time of the first uploading; is the uploading round;

the CPU collects the model parameter abstract sets uploaded by a plurality of edge devices Then, generating global weights by adopting an aggregation method in federal learning:

Wherein, the Is a global regression coefficient vector; the total number of the edge devices; is a global bias term; Is the first The number of samples contained by the edge devices; Represent the first Bias terms for the edge devices; Represent the first Regression coefficients for the individual edge devices;

Generated global weights And the HTTPS or MQTT encrypted by TLS is transmitted to each edge device, and meanwhile, an integrity check code is attached, so that the edge device needs to verify after receiving.

Further, the edge device receives the global weight issued by the central processing unitThen, sequentially executing compatibility check and similarity check;

The compatibility verification ensures that the local model is consistent with the global weight structure by checking the feature dimension and the model version:

Wherein, the Is a compatibility judgment result;And The feature dimension and the version number of the local model are sequentially corresponding; for the indication function, the condition is satisfied with a value of 1, otherwise, the condition is not satisfied with a value of 0; Is a model version number; is a logical operator, represents and;

If it is If the update of the current round is aborted and reported that the update failure causes that the compatibility check fails, ifRepresenting that the compatibility check passes.

Further, after the compatibility verification is passed, whether the weight directions of the local model and the global model are consistent is further compared, and cosine similarity calculation is adopted:

Wherein, the Is a cosine similarity value, and;Is the inner product of internal volume; Is a two-norm;

the cosine similarity degree value to be given And set a threshold valueComparing ifIf the two directions are consistent, the safe update is carried out, otherwise, ifThe fact that the difference between the local weight and the global weight is large is indicated, the local weight is not updated at the moment, the local weight is kept unchanged, and the reason that the local weight is not updated is reported to be insufficient in similarity;

According to the search matching result, the edge device updates the local weight in a direct coverage mode:

,。

Compared with the prior art, the invention has the advantages that:

1. the invention integrates multi-source process data into a structured parameter data file through a modularized link of acquisition, digitalization, encryption storage and hierarchical management, and carries out hierarchical storage according to importance, thereby ensuring that key data can be quickly used, and considering capacity and durability requirements, forming an auditable, traceable and versionable process data base, and further obviously improving the system response and data management quality.

2. According to the invention, the normalized process characteristics are subjected to ridge regression modeling by means of the local model training module, the model parameter abstract is output, and the steady mapping of parameter-quality output is established in continuous iteration, so that the equipment can keep the dynamic adjustability and stable convergence of parameters along with the change of the sensing signals and the running state, and the consistency of printing quality and the process stability are further improved.

3. According to the invention, the global weight is generated by combining federation aggregation through the interaction module, and the compatibility verification and cosine similarity threshold gating of feature dimension/version are introduced at the issuing side, so that a closed loop mechanism of 'safe distribution-judgment update-steady-state coverage' is realized by matching with TLS/MQTT and an integrity verification code, the maintenance cost and mismatch risk are obviously reduced, and the effectiveness and usability of model update are ensured.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic diagram of the system workflow of the present invention;

FIG. 2 is a schematic diagram of a parameter data collection process according to the present invention;

fig. 3 is a schematic diagram of a local weight update flow according to the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present application more apparent, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application, and it is apparent that the described embodiments are some embodiments of the present application, but not all embodiments of the present application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.

In order to achieve the above object, the present invention provides an edge computing storage device for self-learning of printing process parameters, as shown in fig. 1 to 3, the system includes:

and M1, a local acquisition and storage module, which consists of a sensor, a data acquisition board and a flash memory chip, wherein the output signal of the sensor is integrated into a parameter data file by a central processing unit after being digitally processed by the data acquisition board and is stored in the flash memory chip in an encrypted manner.

The sensor comprises a temperature sensor, a humidity sensor, a pressure sensor and a speed sensor, wherein the temperature sensor is used for monitoring the surface temperature and the environment temperature of the printing head and the substrate, the humidity sensor is used for monitoring the humidity of the environment and the substrate, the pressure sensor is used for monitoring the pressure of the printing head, and the speed sensor is used for monitoring the printing speed of the printer;

the detailed data are organized into tables expressed as:

Wherein, the Representing a time round of data acquisition; representing data collected by the corresponding sensor;

the analog signal after sampling Conversion to digital signals by ADCThe conversion process formula is as followsWherein, the method comprises the steps of,AndThe input voltage range of the ADC is sequentially corresponding to the maximum voltage value and the minimum voltage value allowed by the ADC input; the number of bits of the ADC is used for determining the precision of the digital signal;

Based on the analog signal And the digital signalCalculating quantization errorThe formula is;

Setting the maximum quantization error asAnd (2) andIf (1)The original ADC bit number is maintainedUnchanged, ifThen the ADC bit number needs to be increasedTo reduce the error, the calculation formula is:

Wherein, the Is the minimum number of bits of the ADC;

。

and M2, a hierarchical data management module, wherein the central processing unit, the flash memory chip and the EEPROM work cooperatively, the central processing unit classifies importance grades of the parameter data files according to a set period, and the parameter data files are respectively stored in the flash memory chip or the EEPROM after threshold comparison.

Based on the parameter data fileCalculating the variation amplitudeThe formula is:

in the present embodiment, the weight coefficient The value of (2) is 0.7, the weight coefficientIs 0.3;

Wherein, the storage position corresponding to 1 is a flash memory chip, and the storage position corresponding to 2 is an EEPROM;

In the present embodiment, the threshold value The statistical distribution of the history experimental data is analyzed and determined, and the value is 0.75.

And M3, a local model training module acquires the parameter data file, invokes the central processing unit to perform model training on the parameter data file, generates a model parameter abstract, and synchronously writes the model parameter abstract into the EEPROM.

Calculating the mean value and standard deviation of each feature based on the parameter data in the parameter data file:

Wherein, the Is the firstThe mean value of the individual features; Is the first Sample numberInputting characteristic values; Is the number of valid samples; Is the first Standard deviation of individual features; is the feature number;

Composition vector:

Wherein, the Representing dimensions asEach element belonging to a real set;

And then carrying out standardization processing on the parameter data file, wherein the formula is as follows:

Wherein, the Is a normalized feature matrix with the size of;Is an original input matrix with the size of;Is of the size ofIs the whole of (2)Vector, is used for calculating the mean value; to prevent the extremely small positive number of zero removal, and ;

The central processing unit carries out model training on the standardized parameter data in the parameter data file through ridge regression, and the training process is expressed as follows:

Wherein, the Representing a minimization loss function, representing a model training objective to find the best regression coefficient with the minimization loss functionBias term;Is regression coefficient vector with the size ofRepresenting the contribution of each input feature to the target output; is a bias term of the size The method is used for adjusting the reference value output by the model; the regularization parameters are used for controlling the complexity of regression coefficients and preventing overfitting; For the target output vector, the size is Representing the actual output of the printing process; the square representing the euclidean norm, the sum of squares representing the vector, for measuring the error; Is an offset term For expanding to a target output vectorVector of consistent dimensions facilitating correlation in a loss functionSubtracting;

in the present embodiment, the regularization parameters The number of (2) is 0.1;

Wherein, the Is a residual vector of the size ofRepresenting the difference between the predicted value and the actual value; Representing the bias term Extended to and predicted output vectorColumn vectors of consistent dimensions, thereby ensuring that both can be added element by element;

Wherein, the The mean square error is used for measuring the difference between the predicted value and the true value of the model, and the smaller the value is, the higher the model precision is; For determining coefficients, the coefficients are used for measuring the model fitting goodness, and the closer the value is to 1, the better the model fitting effect is;

Generating the model parameter abstract Expressed as:

The edge equipment sets time intervalsUploading the model parameter abstract in the EEPROM to a central processing unit, and calculating the time interval formula asWherein, the method comprises the steps of,The starting time of the first uploading; is the uploading round;

Generated global weights The HTTPS or MQTT protocol encrypted by TLS is issued to each edge device, and meanwhile, an integrity check code (such as SHA 256) is attached, and the edge device needs to verify after receiving;

the edge device receives the global weight issued by the central processing unit Then, sequentially executing compatibility check and similarity check;

If it is The update of the round is stopped and the reason that the update is not reported is that the compatibility check is not passed;

If it is The representative compatibility passes the verification, and further compares whether the weight directions of the local model and the global model are consistent, and cosine similarity calculation is adopted:

In the present embodiment, the threshold value The grid search and ROC/PR analysis of different candidate thresholds are combined with comprehensive evaluation of service indexes to determine that the numerical value is 0.85;

,。

The foregoing is merely illustrative of the present application, and the present application is not limited thereto, and any person skilled in the art will readily recognize that variations or substitutions are within the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Finally, the foregoing description of the preferred embodiment of the invention is not intended to be limiting, but rather to enable any modification, equivalent or improvement to be made, which is intended to be included within the spirit and principles of the invention.

Claims

1. An edge computing storage device for self-learning of printing process parameters, comprising:

The interaction module uploads the model parameter abstract in the EEPROM to the central processing unit in a set time interval, and simultaneously receives global model weights issued by the central processing unit, and the central processing unit searches and matches the global model weights and then covers local model weights;

the analog signal after sampling Conversion to digital signals by ADC;

Wherein, the Is the minimum number of bits of the ADC;

。

2. The printing process parameter self-learning edge calculation storage device of claim 1, wherein the parameter data file is based on Calculating the variation amplitudeThe formula is:

Wherein, the The corresponding storage position is a flash memory chip; The corresponding storage location is EEPROM.

3. The printing process parameter self-learning edge calculation storage device of claim 1 wherein the mean value of each feature is calculated based on parameter data within the parameter data fileAnd standard deviationComposition vector:

Wherein, the Representing dimensions asEach element belonging to a real set;

Generating the model parameter abstract Expressed as:

4. A printing process parameter self-learning edge calculation storage device as claimed in claim 3 wherein the edge device is arranged toUploading the model parameter abstract in the EEPROM to a central processing unit, and calculating the time interval formula asWherein, the method comprises the steps of,The starting time of the first uploading; is the uploading round;

5. The printing process parameter self-learning edge calculation storage device of claim 4, wherein said edge device receives said global weights issued by said central processorThen, sequentially executing compatibility check and similarity check;

6. The printing process parameter self-learning edge computing storage device according to claim 5, wherein after the compatibility verification is passed, further comparing whether the weight directions of the local model and the global model are consistent, and adopting cosine similarity to calculate:

,。