Edge computing storage device for printing process parameter self-learning

Abstract

The application belongs to the technical field of printing equipment optimization and edge computing, and specifically relates to an edge computing storage device for printing process parameter self-learning. First, the application provides an edge computing-based system, which realizes efficient data management and storage through digital processing and encrypted storage of data collected by sensors. Second, local model training and federated learning technology are used to ensure real-time adaptation of dynamic optimization and model updating of the printing process. Finally, a secure interaction module is used to realize intelligent management and efficient operation and maintenance of the printing equipment through global weight updating and intelligent adjustment of the local model. Through the above technical solutions, the application significantly improves the stability and quality consistency of the printing process, reduces manual intervention, and improves the response speed and data governance capability of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of printing equipment optimization and edge computing, and particularly to an edge computing storage device for printing process parameter self-learning. BACKGROUND

[0002] In the running process of the printing equipment, a large amount of process-related data will be generated, including temperature, humidity, pressure, speed and other physical quantities. These data not only reflect the working state of the printing equipment, but also directly affect the precision and quality of the finished product. In order to effectively manage the process, these data need to be stored, processed and analyzed, and saved together with model parameters and other information, so as to provide support for subsequent process adjustment, model training and quality traceability. Common storage media include solid state disks and memory cards, which can realize efficient data recording and management at the device end, guarantee the information integrity and security in the running process, and provide a solid foundation for the optimization and intelligent development of printing process through systematic storage and utilization of process data.

[0003] However, the data storage and processing in the prior art mainly rely on centralized servers, resulting in high data transmission delay. Especially in the industrial printing scene with strict real-time requirements, it cannot meet the demand of low delay and high efficiency. Centralized processing makes the devices far away from the server have strong dependence on the network, which is easily affected by network problems or communication interruption. Secondly, the existing printing equipment usually relies on manual adjustment or preset fixed parameters, so that the process adjustment in the printing process is rough, and cannot automatically adapt to material changes, environmental fluctuations or equipment aging, resulting in low production efficiency and poor product consistency. Finally, the model update in the prior art mainly relies on manual intervention or frequent manual calibration, which not only increases the operation complexity, but also may cause the model to be outdated or not meet the actual running situation, and cannot realize flexible and real-time adaptation and optimization.

[0004] Therefore, the present application proposes an edge computing storage device for printing process parameter self-learning. SUMMARY

[0005] In order to solve the problems of low efficiency of data processing and storage, poor self-adaptability of process parameters and complex model updating and adaptation of the existing printing storage device technology mentioned in the background art, the purpose of the present application is to provide an edge computing storage device for printing process parameter self-learning.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0007] The edge computing storage device for printing process parameter self-learning comprises:

[0008] M1: local acquisition storage module, composed of a sensor, a data acquisition board and a flash chip, the sensor output signal is digitized by the data acquisition board, then integrated by the central processing unit as parameter data file and stored in the flash chip after encryption;

[0009] M2: hierarchical data management module, composed of the central processing unit, the flash chip and EEPROM, the central processing unit divides the parameter data file into importance levels according to the set period, and after threshold comparison, the parameter data file is stored in the flash chip or the EEPROM;

[0010] M3: local model training module, acquires the parameter data file, and calls the central processing unit to train the model of the parameter data file, generates a model parameter summary, and synchronously writes the model parameter summary into the EEPROM;

[0011] M4: interaction module, uploads the model parameter summary in the EEPROM to the central processing unit at a set time interval, and receives the global model weight issued by the central processing unit, and the central processing unit retrieves and matches the global model weight to overwrite the local model weight.

[0012] Further, the sensor includes a temperature sensor, a humidity sensor, a pressure sensor and a speed sensor; the temperature sensor is used to monitor the temperature of the print head, the surface temperature of the substrate and the ambient temperature; the humidity sensor is used to monitor the humidity of the environment and the substrate; the pressure sensor is used to monitor the pressure of the print head; the speed sensor is used to monitor the printing speed of the printer;

[0013] Firstly, the sensor converts the physical quantity collected into an analog electrical signal Output; then, the data acquisition board sets a fixed sampling period , and periodically samples from the analog signal in the period, the sampling process is as follows:

[0014] At time, the sampled analog signal is recorded as , which represents the signal value of the analog electrical signal at time, and ; wherein ; is the initial time; is the sampling index, which represents different sampling times;

[0015] The sampled analog signal is converted into a digital signal by ADC;

[0016] Based on the analog signal and the digital signal , calculate the quantization error , the formula is , wherein, is the maximum voltage value allowed by the ADC input, is the minimum voltage value allowed by the ADC input;

[0017] Set the maximum quantization error as , and ; if , keep the original ADC bit number unchanged; if , increase the ADC bit number to reduce the error, the calculation formula is:

[0018]

[0019] , wherein, is the minimum bit number of the ADC;

[0020] All the digital signals are integrated into parameter data files according to the time sequence , the formula is:

[0021] .

[0022] Further, based on the parameter data files , calculate the change amplitude , the formula is:

[0023]

[0024] Then, based on the change amplitude , calculate the importance level of the first parameter data file , the formula is:

[0025]

[0026] , wherein, is the data source, indicating different types of sensor data; and are weight coefficients, corresponding to the influence degree of the change amplitude and the data source on the importance level, respectively;

[0027] Compare the importance level with a set threshold , and determine the storage location of the parameter data file as follows:

[0028]

[0029] Wherein, 1 corresponds to the storage location of the flash memory chip; 2 corresponds to the storage location of the EEPROM.

[0030] Further, based on the parameter data in the parameter data file, the mean value of each feature is calculated And the standard deviation The composition vector is:

[0031]

[0032] Wherein, Indicates a column vector with a dimension of Each element belongs to the real number set ;

[0033] Then, the parameter data file is standardized; the central processing unit trains the parameter data in the standardized parameter data file through ridge regression, and the training process is represented as:

[0034]

[0035] Wherein, Represents the minimization of the loss function, indicating the goal of model training; The regression coefficient vector is , indicating the contribution of each input feature to the target output; The bias term is , used to adjust the baseline value of the model output; The regularization parameter is used to control the complexity of the regression coefficient to prevent overfitting; The target output vector is , used to represent the actual output of the printing process; Indicates the square of the Euclidean norm, indicating the square sum of the vector, used to measure the error; The standardized feature matrix is The extended form of the bias term is used to extend it to a vector consistent with the target output vector ;

[0036] Based on the minimization of the loss function The estimated value of the regression coefficient is calculated, and the formula is:

[0037]

[0038] Wherein, The estimated value of the regression coefficient is ; is an identity matrix with size ; is an input feature matrix; is an input feature matrix is a transpose matrix of

[0039] Based on the minimized loss function , the estimated value of the bias term is calculated by the formula:

[0040]

[0041] wherein, is the estimated value of the bias term; is the mean of the target output with size ;

[0042] After the training is completed, the training residual error is calculated by the following formula:

[0043]

[0044] wherein, is a residual error vector with size , representing the difference between the predicted value and the true value; represents the expansion of the bias term into a column vector with the same dimension as the predicted output vector ;

[0045] The performance of the ridge regression model is then evaluated using the following criteria:

[0046]

[0047] wherein, is the mean square error, used to measure the difference between the predicted value of the model and the true value; is the coefficient of determination, used to measure the goodness of fit of the model;

[0048] The model parameter summary is generated and represented as:

[0049]

[0050] Finally, the model parameter summary is written synchronously into the EEPROM for storage.

[0051] Further, the edge device uploads the model parameter summary in the EEPROM to the central processor within a set time interval , which is calculated by the formula ; wherein, is the starting time of the first upload; for uploading the round;

[0052] The central processor collects a plurality of model parameter summary sets uploaded by the edge devices After that, the aggregation method in federated learning is used to generate global weights:

[0053]

[0054] wherein, is a global regression coefficient vector; is the total number of edge devices; is a global bias term; is the number of samples contained by the th edge device; represents the bias term of the th edge device; represents the regression coefficient of the th edge device;

[0055] The generated global weights are issued to each edge device through HTTPS or MQTT protocol encrypted by TLS, while an integrity check code is attached, and the edge device needs to verify after receiving.

[0056] Further, the edge device receives the global weights issued by the central processor After that, compatibility verification and similarity test are performed in turn;

[0057] The compatibility verification ensures that the local model and the global weight structure are consistent by checking the feature dimension and the model version:

[0058]

[0059] wherein, is the compatibility determination result; and correspond to the feature dimension and version number of the local model in turn; is an indicator function, which takes the value 1 if the condition is true, and 0 otherwise; is the model version number; is a logical operator, representing and;

[0060] If , the current update is aborted and the non-update reason is reported as the compatibility verification fails; if , it means that the compatibility verification is passed.

[0061] Further, after the compatibility verification is passed, it is further compared whether the weight direction of the local model and the global model is consistent, and the cosine similarity is calculated:

[0062]

[0063] wherein, is a cosine similarity value, and ; is an inner product of inner measure; is a two-norm;

[0064] The cosine similarity value given is compared with a set threshold value If , it is considered that the two directions are consistent, and the safe update is carried out; otherwise, if , it indicates that the local weight and the global weight are greatly different, at this time, the update is not carried out, the local weight is kept unchanged, and the reason for not updating is reported as insufficient similarity;

[0065] According to the search matching result, the edge device updates the local weight in a direct overlay manner:

[0066] , .

[0067] Compared with the prior art, the advantages of the present application are:

[0068] 1. The present application integrates multi-source process data into structured parameter data files and stores them in layers according to importance through the modular link of "collection-digitalization-encrypted storage-hierarchical management", which not only ensures that critical data can be quickly called, but also takes into account the capacity and durability requirements, forms an auditable, traceable and versionable process data base, thereby significantly improving system response and data governance quality.

[0069] 2. The present application outputs model parameter abstracts by means of the local model training module to model the standardized process characteristics through ridge regression, and establishes a robust mapping of parameter-quality output in continuous iteration, so that the device can maintain dynamic adjustment and stable convergence of parameters with the change of sensing signals and operating states, thereby improving printing quality consistency and process stability.

[0070] 3. The present application generates global weights by means of the interaction module combined with federal aggregation, and introduces compatibility verification of feature dimension / version and cosine similarity threshold gating on the side of issuing, cooperates with TLS / MQTT and integrity check code, realizes the closed-loop mechanism of "safe distribution-determination update-steady state coverage", significantly reduces maintenance overhead and mismatch risk, and ensures the effectiveness and availability of model update. BRIEF DESCRIPTION OF DRAWINGS

[0071] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0072] Figure 1 The system workflow schematic diagram of the present application;

[0073] Figure 2 The parameter data collection process schematic diagram of the present application;

[0074] Figure 3 The local weight update process schematic diagram of the present application. DETAILED DESCRIPTION

[0075] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort are within the scope of protection of the present application.

[0076] In order to achieve the above object, the present application is implemented by the following technical solutions. The present application provides an edge computing storage device for self-learning of printing process parameters, as shown in the figure, the system comprises: Figures 1-3

[0077] M1: local acquisition and storage module, composed of a sensor, a data acquisition board and a flash memory chip, the sensor output signal is digitized by the data acquisition board, and then integrated by a central processing unit into a parameter data file and stored in the flash memory chip.

[0078] The sensor comprises a temperature sensor, a humidity sensor, a pressure sensor and a speed sensor; the temperature sensor is used to monitor the temperature of the print head, the surface temperature of the substrate and the ambient temperature; the humidity sensor is used to monitor the humidity of the environment and the substrate; the pressure sensor is used to monitor the pressure of the print head; the speed sensor is used to monitor the printing speed of the printer;

[0079] The detailed data is arranged in a table as follows:

[0080]

[0081] Among them, represents the time round of data acquisition; represents the data collected by the corresponding sensor; ​

[0082] First, the sensor converts the collected physical quantities into analog electrical signals. Output; subsequently, the data acquisition board is set to a fixed sampling period. And within this period, it periodically samples from the analog signal, the sampling process being as follows:

[0083] exist Sample at constant intervals, and record the sampled analog signal as... This indicates that the analog electrical signal is in The signal value at time, and ;in, ; This is the initial time; This is the sampling index, representing different sampling times;

[0084] The sampled analog signal Converted to digital signal by ADC The conversion process formula is as follows ;in, and The input voltage range of the ADC corresponds to the maximum and minimum allowable input voltage values ​​of the ADC, respectively. The number of bits in the ADC determines the precision of the digital signal.

[0085] Based on the analog signal and the digital signal Calculate quantization error The formula is ;

[0086] Set the maximum quantization error to ,and ;like Then maintain the original ADC bit width. Unchanged; if Then the number of ADC bits needs to be increased. To reduce errors, the calculation formula is as follows:

[0087]

[0088] in, This represents the minimum number of bits in the ADC.

[0089] All of the aforementioned digital signals They are integrated into parameter data files according to time sequence. The formula is expressed as:

[0090] .

[0091] M2: hierarchical data management module, cooperated by the central processor, the flash memory chip and EEPROM, the central processor divides the parameter data files into importance levels according to a set period, and after threshold comparison, the parameter data files are respectively stored in the flash memory chip or the EEPROM.

[0092] Based on the parameter data files Calculate the change amplitude , the formula is:

[0093]

[0094] Then, based on the change amplitude Calculate the importance level of the first Parameter data file , the formula is:

[0095]

[0096] Wherein, Data source, representing different types of sensor data; And Both are weight coefficients, corresponding to the influence degree of change amplitude and data source on importance level respectively;

[0097] In this embodiment, the value of the weight coefficient is 0.7; the value of the weight coefficient is 0.3;

[0098] The importance level Is compared with the set threshold , and the storage location of the parameter data file is determined as follows:

[0099]

[0100] Wherein, 1 corresponds to the storage location of the flash memory chip; 2 corresponds to the storage location of the EEPROM;

[0101] In this embodiment, the value of the threshold Is determined by analyzing the statistical distribution of historical experimental data, and is 0.75.

[0102] M3: local model training module, obtains the parameter data file, and calls the central processor to train the parameter data file to generate model parameter summary, and synchronously writes the model parameter summary into the EEPROM.

[0103] Based on the parameter data in the parameter data file, the mean and standard deviation of each feature are calculated:

[0104]

[0105] in, For the first The mean of each feature; For the first The first sample One input feature value; The number of valid samples; For the first The standard deviation of each feature; The characteristic number;

[0106] Composition vector:

[0107]

[0108] in, The dimension is A column vector, where each element belongs to the set of real numbers. ;

[0109] The parameter data file is then standardized using the following formula:

[0110]

[0111] in, The standardized feature matrix has a size of ; The original input matrix has a size of . ; It is the size of The whole Vectors, used to calculate the mean; To prevent extremely small positive numbers from being divided by zero, and ;

[0112] The central processing unit trains a model on the standardized parameter data in the parameter data file using ridge regression. The training process is as follows:

[0113]

[0114] in, This represents minimizing the loss function, indicating the objective of model training: to find the optimal regression coefficients by minimizing the loss function. and bias terms ; This is a vector of regression coefficients, with a size of... , representing the contribution of each input feature to the target output; This is the bias term, with a size of , used to adjust the baseline value of the model output; This is a regularization parameter used to control the complexity of the regression coefficients and prevent overfitting. The target output vector has a size of This is used to represent the actual output of the printing process; The square of the Euclidean norm is represented by the sum of squares of a vector, which is used to measure error. For bias terms The extended form is used to expand it to the target output vector. Vectors with consistent dimensions are easier to integrate with the loss function. Subtraction;

[0115] In this embodiment, the regularization parameter The value is 0.1;

[0116] Based on the minimized loss function Calculate regression coefficients The estimated value is given by the formula:

[0117]

[0118] in, These are estimates of the regression coefficients, with a magnitude of... ; It is an identity matrix with a size of ; The input feature matrix; Input feature matrix The transpose of the matrix;

[0119] Based on the minimized loss function Calculate the bias term The estimated value is given by the formula:

[0120]

[0121] in, This is an estimate of the bias term; The mean of the target output, with a magnitude of ;

[0122] After training is complete, the training residuals are calculated using the following formula:

[0123]

[0124] in, Let be the residual vector, with a size of This represents the difference between the predicted value and the actual value; Indicates the bias term Expanded to the predicted output vector Column vectors with the same dimensions ensure that the two can be added element by element;

[0125] The performance of the ridge regression model was then evaluated using the following criteria:

[0126]

[0127] in, Mean squared error is used to measure the difference between the model's predicted values ​​and the actual values. The smaller the value, the higher the model's accuracy. The coefficient of determination is used to measure the goodness of fit of the model. The closer the value is to 1, the better the model fit.

[0128] Generate the model parameter summary , represented as:

[0129]

[0130] Finally, the model parameter summary Synchronously write to EEPROM for storage.

[0131] M4: Interaction module, which uploads the model parameter summary in the EEPROM to the central processing unit within a set time interval, and simultaneously receives the global model weights issued by the central processing unit. After the central processing unit searches for a match, it overwrites the local model weights with the global model weights.

[0132] Edge devices at set time intervals Within the process, the model parameter summary in the EEPROM is uploaded to the central processing unit, and the time interval is calculated using the following formula: ;in, The start time of the first upload; For the upload round;

[0133] The central processing unit collects model parameter summary sets uploaded by multiple edge devices. Then, the aggregation method in federated learning is used to generate global weights:

[0134]

[0135] in, This is the global regression coefficient vector; Total number of edge devices; This is a global bias term; For the first The number of samples contained in each edge device; Indicates the first The bias term for each edge device; Indicates the first Regression coefficients of each edge device;

[0136] Generated global weights The data is distributed to each edge device via TLS-encrypted HTTPS or MQTT protocol, along with an integrity checksum (e.g., ...). SHA256 The edge device needs to verify the data after receiving it.

[0137] The edge device receives the global weight issued by the central processing unit. Then, compatibility checks and similarity checks are performed sequentially.

[0138] The compatibility check ensures that the local model and the global weight structure are consistent by checking the feature dimensions and model version.

[0139]

[0140] in, This is the compatibility assessment result; and These correspond to the feature dimensions and version number of the local model, respectively. This is an indicator function; it takes the value 1 if the condition is true, and 0 if the condition is false. This refers to the model version number; For logical operators, AND is used;

[0141] like If so, this round of updates will be suspended and the reason for not updating will be reported as a failure of compatibility verification;

[0142] like This indicates that the compatibility check has passed. Further comparison is made to determine if the weight directions of the local and global models are consistent, using cosine similarity calculation.

[0143]

[0144] in, Let be the cosine similarity value, and ; Inner product of inner quantities; It is a norm 2;

[0145] The given cosine similarity values With set threshold If a comparison is made, If both directions are aligned, a security update will be performed; otherwise, if... This indicates a large difference between the local weight and the global weight. In this case, no update is performed, the local weight remains unchanged, and the reason for not updating is reported as insufficient similarity.

[0146] In this embodiment, the threshold Through grid search and ROC / PR analysis of different candidate thresholds, combined with a comprehensive evaluation of business indicators, the value was determined to be 0.85;

[0147] Based on the search matching results, edge devices update their local weights using a direct overwrite method:

[0148] , .

[0149] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0150] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An edge computing storage device for self-learning printing process parameters, characterized in that, include: M1: Local data acquisition and storage module, consisting of a sensor, a data acquisition board and a flash memory chip. The sensor output signal is digitized by the data acquisition board and then integrated by the central processing unit into a parameter data file and stored in the flash memory chip in an encrypted manner. M2: Hierarchical data management module, which works in concert with the central processing unit, the flash memory chip and the EEPROM. The central processing unit classifies the parameter data files into importance levels according to a set period, and after threshold comparison, stores the parameter data files into the flash memory chip or the EEPROM respectively. M3: Local model training module, which acquires the parameter data file, calls the central processing unit to train the model on the parameter data file, generates a model parameter summary, and synchronously writes the model parameter summary into the EEPROM; M4: Interaction module, which uploads the model parameter summary in the EEPROM to the central processing unit within a set time interval, and simultaneously receives the global model weight issued by the central processing unit. After the central processing unit searches for a match, it overwrites the local model weight with the global model weight. The sensors include a temperature sensor, a humidity sensor, a pressure sensor, and a speed sensor; the temperature sensor is used to monitor the temperature of the printhead, the substrate surface, and the ambient temperature; the humidity sensor is used to monitor the humidity of the environment and the substrate; the pressure sensor is used to monitor the pressure of the printhead; and the speed sensor is used to monitor the printing speed of the printer. First, the sensor converts the collected physical quantities into analog electrical signals. Output; subsequently, the data acquisition board is set to a fixed sampling period. And within this period, it periodically samples from the analog signal, the sampling process being as follows: exist Sample at constant intervals, and record the sampled analog signal as... This indicates that the analog electrical signal is in The signal value at time, and ;in, ; This is the initial time; This is the sampling index, representing different sampling times; The sampled analog signal Converted to digital signal by ADC ; Based on the analog signal and the digital signal Calculate quantization error The formula is ,in, Set the maximum allowable voltage value for the ADC input. Set the minimum allowable voltage value for the ADC input; Set the maximum quantization error to ,and ;like Then maintain the original ADC bit width. Unchanged; if Then the number of ADC bits needs to be increased. To reduce errors, the calculation formula is as follows: in, This represents the minimum number of bits in the ADC. All of the aforementioned digital signals They are integrated into parameter data files according to time sequence. The formula is expressed as: 。 2. The edge computing storage device for self-learning printing process parameters according to claim 1, characterized in that, Based on the parameter data file Calculate the range of change The formula is: Subsequently, based on the magnitude of the change Calculation yields the first The importance level of the parameter data files The formula is: in, The data source indicates different types of sensor data; and These are all weighting coefficients, corresponding to the magnitude of change and the degree of influence of the data source on the importance level, respectively; The importance level With set threshold The storage location of the parameter data file is determined by comparison as follows: in, The corresponding storage location is a flash memory chip; The corresponding storage location is EEPROM.

3. The edge computing storage device for self-learning printing process parameters according to claim 1, characterized in that, Based on the parameter data in the parameter data file, calculate the mean value of each feature. and standard deviation Composition vector: in, The dimension is A column vector, where each element belongs to the set of real numbers. ; The parameter data file is then standardized; the central processing unit trains a model on the standardized parameter data file using ridge regression. The training process is as follows: in, This represents minimizing the loss function, indicating the objective of model training; This is a vector of regression coefficients, with a size of... , representing the contribution of each input feature to the target output; This is the bias term, with a size of , used to adjust the baseline value of the model output; This is a regularization parameter used to control the complexity of the regression coefficients and prevent overfitting. The target output vector has a size of This is used to represent the actual output of the printing process; The square of the Euclidean norm is represented by the sum of squares of a vector, which is used to measure error. The standardized feature matrix; For bias terms The extended form is used to expand it to the target output vector. Vectors with consistent dimensions; Based on the minimized loss function Calculate regression coefficients The estimated value is given by the formula: in, These are estimates of the regression coefficients, with a magnitude of... ; It is an identity matrix with a size of ; The input feature matrix; Input feature matrix The transpose of the matrix; Based on the minimized loss function Calculate the bias term The estimated value is given by the formula: in, This is an estimate of the bias term; The mean of the target output, with a magnitude of ; After training is complete, the training residuals are calculated using the following formula: in, Let be the residual vector, with a size of This represents the difference between the predicted value and the actual value; Indicates the bias term Expanded to the predicted output vector A column vector with consistent dimensions; The performance of the ridge regression model was then evaluated using the following criteria: in, Mean squared error is used to measure the difference between the model's predicted value and the actual value. The coefficient of determination is used to measure the goodness of fit of the model. Generate the model parameter summary , is represented as: Finally, the model parameter summary Synchronously write to EEPROM for storage.

4. The edge computing storage device for self-learning printing process parameters according to claim 3, characterized in that, Edge devices at set time intervals Within this process, the model parameter summary in the EEPROM is uploaded to the central processing unit, and the time interval is calculated using the following formula: ;in, The start time of the first upload; For the upload round; The central processing unit collects model parameter summary sets uploaded by multiple edge devices. Then, the aggregation method in federated learning is used to generate global weights: in, This is the global regression coefficient vector; Total number of edge devices; This is a global bias term; For the first The number of samples contained in each edge device; Indicates the first The bias term for each edge device; Indicates the first Regression coefficients of each edge device; Generated global weights The message is sent to each edge device via TLS-encrypted HTTPS or MQTT protocol, along with an integrity checksum, which the edge device must verify upon receipt.

5. The edge computing storage device for self-learning printing process parameters according to claim 4, characterized in that, The edge device receives the global weight issued by the central processing unit. Then, compatibility checks and similarity checks are performed sequentially. The compatibility check ensures that the local model and the global weight structure are consistent by checking the feature dimensions and model version. in, This is the compatibility assessment result; and These correspond to the feature dimensions and version number of the local model, respectively. This is an indicator function; it takes the value 1 if the condition is true, and 0 if the condition is false. This refers to the model version number; For logical operators, AND is used; like If the update is not completed, this round of updates will be suspended and the reason for not updating will be reported as a failure to pass compatibility verification; if This indicates that the compatibility check has passed.

6. The edge computing storage device for self-learning printing process parameters according to claim 5, characterized in that, After the compatibility check is passed, the weight directions of the local model and the global model are further compared to see if they are consistent, and cosine similarity is used for calculation. in, Let be the cosine similarity value, and ; Inner product of inner quantities; It is a norm 2; The given cosine similarity values With set threshold If a comparison is made, If both directions are aligned, a security update will be performed; otherwise, if... This indicates a large difference between the local weight and the global weight. In this case, no update is performed, the local weight remains unchanged, and the reason for not updating is reported as insufficient similarity. Based on the search matching results, edge devices update their local weights using a direct overwrite method: ， 。

Images