Data processing method, device and equipment based on artificial intelligence technology and data model

Abstract

The present invention relates to a data processing method, device and equipment based on artificial intelligence technology and data model. The data processing method includes the following steps: obtaining historical process parameter data of original production equipment and real-time process parameter data of new equipment; constructing an association network model reflecting the strength of mutual influence of parameters; performing partial correlation analysis and cross-correlation analysis on the real-time process parameters of new equipment and the historical process parameters of original production equipment; incrementally updating the process parameter prediction model; integrating the real-time process parameter feature representation of the new equipment into the association network model; applying the optimal process parameter configuration set to the actual production process; and continuously accumulating and improving the adaptive knowledge introduced by the new equipment. Through continuous learning and knowledge accumulation, the present invention forms a set of dynamically optimized closed-loop control mechanisms, which can effectively cope with the changes in process parameters brought about by the introduction of new equipment, and realize continuous optimization and quality improvement of the production process.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a data processing method, device and equipment based on artificial intelligence technology and a data model. Background Art

[0002] In the data processing of artificial intelligence technology and data models, when new equipment is introduced into the production process to replace the original equipment, the dynamic correlation between process parameters may change; the introduction of new equipment may bring many changes, such as equipment operating speed, accuracy, production capacity, energy consumption, etc. These changes may affect other related process parameters. For example, when the speed of a certain equipment is increased, the equipment in the subsequent process may need to adjust the feed speed, temperature, pressure and other parameters accordingly to adapt to the new production rhythm. The feature combination method that the original parameter prediction model relies on may no longer be applicable and needs to be adjusted according to the characteristics of the new equipment.

[0003] However, the adjustment of feature combinations may trigger a series of chain reactions, affecting the generation of process optimization suggestions. The key issue here is: how to dynamically adjust the feature combination method of the parameter prediction model to adapt it to the characteristics of the new equipment while ensuring the stability of the process flow; at the same time, how to quantitatively analyze the impact mechanism of feature combination adjustment on process optimization suggestions to ensure the reliability and effectiveness of the optimization suggestions. This requires in-depth research on the differences in process parameter correlation between new and old equipment, the establishment of a dynamic feature selection and combination mechanism for the parameter prediction model, and the construction of an impact mechanism model for process optimization suggestions. It comprehensively considers the multi-dimensional influencing factors of feature combination adjustment to provide accurate and reliable optimization decision support. Summary of the invention

[0004] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a data processing method, device and equipment based on artificial intelligence technology and data model. The data processing method based on artificial intelligence technology and data model forms a set of dynamic optimization closed-loop control mechanisms through continuous learning and knowledge accumulation, which can effectively cope with the changes in process parameters brought about by the introduction of new equipment and realize continuous optimization and quality improvement of the production process.

[0005] The data processing method based on artificial intelligence technology and data model described in the present invention comprises the following steps:

[0006] S1. Obtain the historical process parameter data of the original production equipment and the real-time process parameter data of the new equipment, and perform data preprocessing to obtain a standardized time series process parameter data set;

[0007] S2. By mining the dynamic correlation characteristics between different process parameters and determining the causal relationship between the parameters, a correlation network model reflecting the intensity of mutual influence of the parameters is constructed;

[0008] S3. Conduct partial correlation analysis and cross-correlation analysis on the real-time process parameters of the new equipment and the historical process parameters of the original production equipment, and build a process parameter prediction model;

[0009] S4, incrementally updating the process parameter prediction model, and using an adaptive learning algorithm to adjust the parameters of the process parameter prediction model and the weight of evaluating new samples;

[0010] S5. Through the knowledge increment fusion method, the real-time process parameter feature representation of the new equipment is integrated into the association network model, and the parameter combination is optimized to obtain the optimal process parameter configuration set suitable for the new equipment;

[0011] S6. Applying the optimal process parameter configuration set to the actual production process, adjusting and monitoring the parameters of the new equipment in real time, and dynamically correcting the optimal process parameter configuration set through a feedback mechanism;

[0012] S7. Utilize an incremental learning algorithm to continuously update the association network model and the optimization decision knowledge base, continuously accumulate and improve the adaptive knowledge introduced by the new equipment, and form a dynamic optimization closed-loop control mechanism.

[0013] Preferably, the step S1 specifically includes:

[0014] Based on the historical process parameter data of the original production equipment and the real-time process parameter data of the new equipment, data inspection is performed by setting a time window with a fixed time interval;

[0015] For process parameter data with missing values ​​in parameter records, linear interpolation method is used to obtain supplementary filling data;

[0016] Calculate the mean and standard deviation of the parameter values ​​according to the supplementary filling data, and obtain the outlier elimination result according to whether the parameter exceeds the preset standard deviation interval;

[0017] According to the abnormal value elimination result, a parameter type comparison table of temperature parameters, pressure parameters and flow parameters is established, and dimensionally unified parameters are obtained through Kelvin conversion, Pascal conversion and cubic meters per second conversion;

[0018] The dimensionally unified parameters are normalized to obtain normalized values, and the Pearson correlation coefficient is calculated for the normalized values ​​using a sliding time window to determine the process parameter mapping relationship to obtain a standardized time series process parameter data set.

[0019] Preferably, the step S2 specifically includes:

[0020] According to the process parameter time series data in the time series process parameter data set, a parameter autocorrelation coefficient sequence is calculated using a sliding window autocorrelation function, and the parameter change period is determined by identifying the time delay value corresponding to the first negative value of the autocorrelation coefficient;

[0021] According to the parameter variation cycle, the Granger causality test method is used to calculate the significance level value of the parameter pair. If the significance level value is less than the preset significance threshold, it is determined that there is a causal relationship between the parameter pair;

[0022] For the parameter pairs with causal relationship, the conditional mutual information entropy method is used to calculate the information gain value, and the dynamic correlation strength value of the parameter pair is obtained by comparing the information gain value with a preset information gain threshold;

[0023] According to the dynamic association strength value, a dynamic Bayesian network method is used to construct a parameter influence relationship diagram, and a network connection weight coefficient is established through the directionality of the parameter pairs in the causal relationship table.

[0024] Preferably, the step S3 specifically includes:

[0025] According to the real-time process parameters in the real-time process parameter data, a partial autocorrelation function is used to calculate the autocorrelation coefficient of each parameter in the real-time process parameter sequence;

[0026] An autocorrelation coefficient threshold is set according to the mean and standard deviation of the autocorrelation coefficient, and a difference operation is performed on the real-time process parameter sequence exceeding the autocorrelation coefficient threshold to obtain a stabilized process parameter sequence;

[0027] Calculating the cross-correlation coefficient between the corresponding process parameters of the new equipment and the original production equipment according to the stabilized process parameter sequence, and obtaining a difference process parameter pair if the cross-correlation coefficient is lower than a preset cross-correlation coefficient determination threshold;

[0028] Decomposing the difference process parameter pair by wavelet transform, and obtaining frequency domain eigenvalues ​​by calculating wavelet coefficients at different scales;

[0029] According to the frequency domain eigenvalues, a frequency component combination satisfying a preset contribution rate is selected using a singular spectrum analysis method;

[0030] According to the frequency component combination, a long short-term memory network structure is constructed, and the features of the real-time process parameter sequence are extracted through a sliding time window. The process parameter prediction model is obtained according to the feature and prediction target training.

[0031] Preferably, the step S4 specifically includes:

[0032] An exponential decay function is used to calculate a time weight value of the historical process parameter data, and the historical process parameter data is weighted according to the time weight value to obtain a weighted historical data set;

[0033] Dividing a sliding time window according to the weighted historical data set, and obtaining a data distribution feature vector by calculating the statistics of data samples in the sliding time window;

[0034] Calculating the difference value of the new and old data samples according to the data distribution feature vector, and marking the new and old data samples whose difference value exceeds a preset difference threshold as abnormal samples;

[0035] The residual size of the abnormal sample is calculated, an adaptive learning rate value is determined according to the residual size, parameters of the process parameter prediction model are updated by the adaptive learning rate value, and a prediction result is obtained.

[0036] Preferably, the step S5 specifically includes:

[0037] A deep autoencoder is used to perform dimension reduction encoding on the real-time process parameters, and a compressed feature representation of the real-time process parameters is obtained by setting the number of encoding layer nodes to a preset range of the original parameter dimension;

[0038] Calculating the cosine similarity between feature vectors for the compressed feature representation, merging feature vectors whose cosine similarity is higher than a preset similarity threshold, and obtaining a fused feature vector set;

[0039] Performing a recursive feature elimination operation on the fused feature vector set, eliminating features whose feature importance scores are lower than an importance threshold, and obtaining a key feature combination;

[0040] A sampling point sequence is generated within the historical value interval of the process parameters according to the key feature combination, and an optimal process parameter configuration set suitable for the new equipment is obtained by executing Monte Carlo simulation calculation and multi-objective evaluation function operation.

[0041] Preferably, the step S6 specifically includes:

[0042] A proportional-integral-differential controller is used to obtain a deviation mean and a standard deviation of the real-time process parameter, and a proportional coefficient value is obtained according to the inverse of the standard deviation;

[0043] Collect product quality index data according to the numerical value of the real-time process parameter, and obtain hardness index value, strength index value and purity index value through a data collector;

[0044] Calculating the center line value and the process standard deviation of the quality indicator data, and if the quality indicator data is within the preset standard deviation interval of the center line value, using Fourier transform to obtain the frequency domain characteristic spectrum;

[0045] The frequency domain characteristic spectrum is subjected to three-layer wavelet decomposition, and a quality evaluation score is obtained by calculating the high-frequency band energy value, the mid-frequency band energy value and the low-frequency band energy value, and the control parameters of the process parameters are dynamically adjusted according to the quality evaluation score.

[0046] Preferably, the step S7 specifically includes:

[0047] Calculating an update amount of a parameter association matrix according to process parameters and quality indicator data, wherein the update amount is determined by an incremental gradient descent algorithm and a baseline learning rate;

[0048] According to the parameter association matrix, a sliding time window method is used to obtain the probability distribution of new and old samples, and the chi-square distance value of the sample distribution is calculated;

[0049] According to the chi-square distance value, a process parameter optimizer is constructed by using a random forest algorithm, the process parameter optimizer obtains an optimal split feature by calculating a Gini index, and the structure of the process parameter optimizer is incrementally updated according to the optimal split feature;

[0050] According to the process parameter optimizer, a proportional-integral controller is used to calculate the parameter adjustment amount, and the process parameters are adjusted in real time according to the parameter adjustment amount.

[0051] The present invention also proposes a data processing device based on artificial intelligence technology and a data model, comprising:

[0052] Data collection and preprocessing module, used to collect historical process parameter data of original production equipment and real-time process parameter data of new equipment, and perform data preprocessing;

[0053] Model building module, used to build correlation network model and process parameter prediction model;

[0054] The model optimization and application module is used to integrate the real-time process parameter feature representation of the new equipment into the association network model, obtain the optimal process parameter configuration set suitable for the new equipment, apply the optimal process parameter configuration set to the actual production process, and perform dynamic optimization through the feedback mechanism;

[0055] The knowledge accumulation and closed-loop control module is used to continuously update the associated network model and the optimization decision-making knowledge base, continuously accumulate and improve the adaptive knowledge introduced by new equipment, and form a dynamically optimized closed-loop control mechanism.

[0056] The present invention also proposes a data processing device based on artificial intelligence technology and a data model, comprising:

[0057] at least one memory;

[0058] at least one processor;

[0059] at least one program;

[0060] The program is stored in the memory, and the processor executes at least one of the programs to implement the data processing method based on artificial intelligence technology and data model as described above.

[0061] The data processing method, device and equipment based on artificial intelligence technology and data model described in the present invention have the following advantages:

[0062] The data processing method based on artificial intelligence technology and data model of the present invention obtains the historical process parameter data of the original production equipment and the real-time process parameter data of the new equipment, and performs data preprocessing, thereby providing a reliable basis for subsequent analysis and modeling; by constructing an association network model that reflects the strength of mutual influence of parameters, it is helpful to deeply understand the interaction relationship between process parameters and provide strong support for optimizing process parameter configuration; by constructing a process parameter prediction model, it is possible to accurately predict the process parameters of the new equipment and provide a scientific basis for parameter adjustment in the production process; by integrating the real-time process parameter feature representation of the new equipment into the association network model and optimizing the parameter combination, the optimal process parameter configuration set suitable for the new equipment is obtained, which helps to achieve optimization of the production process and improve efficiency; by using an incremental learning algorithm to continuously update the association network model and the optimization decision knowledge base, continuously accumulate and improve the adaptive knowledge introduced by the new equipment, it can adapt to changes and new situations in the production process and maintain the accuracy and effectiveness of the model; the optimal process parameter configuration set is dynamically corrected through a feedback mechanism to ensure the accuracy and adaptability of the parameter configuration. Through continuous learning and knowledge accumulation, the present invention has formed a set of dynamically optimized closed-loop control mechanisms, which can effectively cope with the changes in process parameters brought about by the introduction of new equipment and achieve continuous optimization and quality improvement of the production process.

[0063] The data processing device based on artificial intelligence technology and data model of the present invention includes a data acquisition and preprocessing module, which is used to acquire historical process parameter data of existing production equipment and real-time process parameter data of new equipment, and perform data preprocessing; a model construction module, which is used to construct an associated network model and a process parameter prediction model; a model optimization and application module, which is used to integrate the real-time process parameter feature representation of the new equipment into the associated network model, obtain the optimal process parameter configuration set suitable for the new equipment, apply the optimal process parameter configuration set to the actual production process, and dynamically optimize it through a feedback mechanism; a knowledge accumulation and closed-loop control module, which is used to continuously update the associated network model and the optimization decision knowledge base, continuously accumulate and improve the adaptive knowledge introduced by the new equipment, and form a dynamically optimized closed-loop control mechanism. The data processing device can quickly and accurately collect and pre-process historical and real-time process parameter data, providing a solid foundation for subsequent model construction; by constructing an associative network model and a process parameter prediction model, it can deeply understand the complex relationships in the production process and predict future process parameter changes; it can integrate the real-time process parameter characteristics of new equipment into the model to achieve rapid adaptation and optimization; and by continuously updating the associative network model and optimizing the decision-making knowledge base, it continuously accumulates and improves adaptive knowledge to form a dynamically optimized closed-loop control mechanism, which not only improves production efficiency and product quality, but also reduces operating costs and equipment failure rates.

[0064] The data processing device based on artificial intelligence technology and data model of the present invention includes at least one memory; at least one processor; at least one program; the program is stored in the memory, and the processor executes at least one program to implement the data processing method based on artificial intelligence technology and data model as described above. The data processing device can process and analyze complex process parameter data in real time through advanced algorithms and models, providing a scientific basis for decision-making in the production process; at the same time, its intelligent optimization algorithm can automatically recommend the optimal process parameter configuration, reducing the risk and error of human decision-making; the high degree of automation and adaptability of the equipment enables it to quickly adapt to the introduction of new equipment and changes in the production process, and maintain the continuity of the optimization effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of the data processing method based on artificial intelligence technology and data model described in the present invention. DETAILED DESCRIPTION

[0066] like Figure 1 As shown, the data processing method based on artificial intelligence technology and data model of the present invention comprises the following steps:

[0067] S1. Obtain the historical process parameter data of the original production equipment and the real-time process parameter data of the new equipment, and perform data preprocessing to obtain a standardized time series process parameter data set;

[0068] S2. By mining the dynamic correlation features between different process parameters and determining the causal relationship between the parameters, a correlation network model reflecting the strength of mutual influence of the parameters is constructed; specifically, mining the dynamic correlation features between different process parameters is achieved by using the time series analysis method;

[0069] S3. Perform partial correlation analysis and cross-correlation analysis on the real-time process parameters of the new equipment and the historical process parameters of the original production equipment, and build a process parameter prediction model; specifically, when introducing new equipment into the production process, obtain the real-time process parameters of the new equipment, perform partial correlation analysis and cross-correlation analysis on the real-time process parameters of the new equipment and the historical process parameters of the original production equipment, and according to the difference between the parameters of the new equipment and the original association network model, use the spectral analysis method to determine the key parameter combination that needs to be adjusted after the introduction of the new equipment, and build a process parameter prediction model;

[0070] S4. Incrementally update the process parameter prediction model, and use an adaptive learning algorithm to adjust the parameters of the process parameter prediction model and the weight of evaluating new samples; specifically, the incremental update of the process parameter prediction model is implemented by an online learning method, and the weight of old data is reduced by a historical data forgetting mechanism, and the parameters of the process parameter prediction model and the weight of evaluating new samples are adjusted by an adaptive learning algorithm, so that the process parameter prediction model can adapt to new parameter distribution characteristics;

[0071] S5. Through the knowledge increment fusion method, the real-time process parameter feature representation of the new equipment is integrated into the association network model, and the parameter combination is optimized to obtain the optimal process parameter configuration set suitable for the new equipment;

[0072] S6. Apply the optimal process parameter configuration set to the actual production process, adjust and monitor the parameters of the new equipment in real time, and dynamically modify the optimal process parameter configuration set through a feedback mechanism;

[0073] S7. Use incremental learning algorithms to continuously update the associated network model and optimization decision knowledge base, continuously accumulate and improve the adaptive knowledge introduced by new equipment, and form a dynamically optimized closed-loop control mechanism.

[0074] Furthermore, in this embodiment, step S1 specifically includes:

[0075] Based on the historical process parameter data of the original production equipment and the real-time process parameter data of the new equipment, data inspection is performed by setting a time window with a fixed time interval;

[0076] For process parameter data with missing values ​​in parameter records, linear interpolation method is used to obtain supplementary filling data;

[0077] The mean and standard deviation of the parameter values ​​are calculated based on the supplementary filling data, and the outlier elimination results are obtained based on whether the parameter exceeds the preset standard deviation range;

[0078] According to the results of outlier elimination, a parameter type comparison table of temperature parameters, pressure parameters and flow parameters is established, and dimensionally unified parameters are obtained through Kelvin conversion, Pascal conversion and cubic meters per second conversion;

[0079] Perform minimum and maximum standardization processing on the dimensionally unified parameters to obtain normalized values, use a sliding time window to calculate the Pearson correlation coefficient of the normalized values, determine the process parameter mapping relationship, and obtain a standardized time series process parameter data set;

[0080] Here is an example:

[0081] Collect historical process parameter data of the original production equipment and perform data integrity checks by setting a fixed time window of ten minutes;

[0082] For process parameter data with missing values ​​in parameter records, a linear interpolation algorithm is used to fill in the missing values;

[0083] For the historical process parameter data after supplementation, set the parameter fluctuation range judgment threshold, calculate the mean and standard deviation of the parameter values, and identify and eliminate the abnormal values ​​that exceed the range of plus or minus three times the standard deviation of the mean;

[0084] Based on the real-time process parameter data collected by the new equipment, the same time window as the historical parameters is used to check the data integrity and supplement the data, and the same parameter fluctuation range determination method is used to eliminate abnormal values;

[0085] For the historical process parameters and real-time process parameters after removing abnormal values, three types of process parameters, namely temperature parameters, pressure parameters and flow parameters, are extracted respectively, and a parameter type comparison table is established;

[0086] According to the established parameter type comparison table, temperature parameters are converted to degrees Kelvin, pressure parameters are converted to Pascal, and flow parameters are converted to cubic meters per second, so as to achieve the dimensional unification of process parameters;

[0087] For the process parameters with unified dimensions, the minimum and maximum value normalization method is used to perform numerical normalization processing, and the normalized value of each process parameter in the corresponding time window is calculated;

[0088] The sliding time window method is used to align the normalized historical process parameters with the real-time process parameters, and the mapping relationship between the corresponding process parameters of the new and old equipment is determined by calculating the Pearson correlation coefficient.

[0089] In the process of processing process parameter data, the selection of fixed time window has an important impact on data integrity. Taking the temperature parameter collection of the production line as an example, the sensor records the temperature value every 1 minute and sets a fixed time window of 10 minutes. If the temperature value at a certain time point is missing, it is supplemented by linear interpolation method. Assuming that the temperature value at 8:20 is missing, the temperature value at 8:19 is 185 degrees Celsius, and the temperature value at 8:21 is 189 degrees Celsius, then the temperature value at 8:20 is calculated as 187 degrees Celsius;

[0090] For the identification and elimination of outliers, the statistical characteristics of the calculation parameters are used for judgment. Taking the pressure parameter as an example, the mean value of the pressure value in a certain period of time is 500 kPa, and the standard deviation is 20 kPa. The positive and negative three times of the standard deviation are set as the fluctuation range judgment threshold, that is, 440 kPa to 560 kPa. The pressure values ​​beyond this range are eliminated to ensure the reliability of the data.

[0091] This method is also used in real-time data processing to maintain consistency in data processing between new and old equipment. The unification of process parameter dimensions is of great significance to data standardization. Taking temperature parameters as an example, the temperature values ​​recorded by different equipment may use different units of measurement, some in Celsius and some in Fahrenheit. They need to be uniformly converted to Kelvin. The calculation formula is Kelvin = Celsius + 273.15. The conversion formula for Fahrenheit is:

[0092] Kelvin = (Fahrenheit + 459.67) * 5 / 9

[0093] The unit conversion is used to unify the dimensions of the temperature parameters. In the parameter normalization process, the minimum and maximum value standardization method is used to map the parameters of different dimensions to the range of 0 to 1. The calculation formula is:

[0094] Normalized value = (current value - minimum value) / (minimum value - maximum value)

[0095] Taking the flow parameter as an example, the maximum flow rate in a certain time window is 100 cubic meters per second, the minimum flow rate is 20 cubic meters per second, and the current value is 60 cubic meters per second. The normalized value is 0.5;

[0096] The establishment of the mapping relationship between the parameters of the new and old equipment is realized by calculating the Pearson correlation coefficient. The correlation coefficient calculation formula is:

[0097]

[0098] in, Indicates historical parameter values; Represents the average value of historical parameters; Indicates real-time parameter value; Indicates the average value of real-time parameters; correlation coefficient The value range of is -1 to 1. When the correlation coefficient is greater than 0.8, it is considered to be strongly correlated, and the mapping relationship between the parameters is determined;

[0099] Taking temperature parameters as an example, the correlation coefficient of the temperature series of the new and old equipment in the same time window was calculated to be 0.92, indicating that the two have a strong correlation and a corresponding mapping relationship can be established.

[0100] Furthermore, in this embodiment, step S2 specifically includes:

[0101] According to the process parameter time series data in the time series process parameter data set, a parameter autocorrelation coefficient sequence is calculated by using a sliding window autocorrelation function, and the parameter change period is determined by identifying the time delay value corresponding to the first negative value of the autocorrelation coefficient;

[0102] According to the parameter change cycle, the Granger causality test method is used to calculate the significance level value of the parameter pair. If the significance level value is less than the preset significance threshold, it is determined that there is a causal relationship between the parameter pairs.

[0103] For parameter pairs with causal relationships, the conditional mutual information entropy method is used to calculate the information gain value. By comparing the information gain value with the preset information gain threshold, the dynamic correlation strength value of the parameter pair is obtained;

[0104] According to the dynamic correlation strength value, the dynamic Bayesian network method is used to construct the parameter influence relationship diagram, and the network connection weight coefficient is established through the directionality of the parameter pairs in the causal relationship table;

[0105] Here is an example:

[0106] According to the process parameter time series data in the time series process parameter data set, the autocorrelation coefficient of each parameter in the time delay sequence is calculated by using the sliding window autocorrelation function, and the time delay value corresponding to the first negative value of the autocorrelation coefficient is identified as the parameter change period, from which the parameter change feature sequence is obtained;

[0107] According to the obtained parameter change characteristic sequence, the Granger causality test method is used to calculate the different process parameters in pairs, and the causal relationship between the parameter pairs is determined by setting the significance level value to 0.05. If the p value is less than the significance level, it is determined that there is a causal relationship between the parameter pairs, and the process parameter causal relationship table is obtained;

[0108] According to the process parameter causal relationship table, the conditional mutual information entropy method is used to calculate the information gain value of the parameter pair. By calculating the joint probability distribution and marginal probability distribution of the parameter pair in different time windows, the dynamic correlation strength value of the parameter pair is obtained.

[0109] According to the dynamic correlation strength value sequence of parameter pairs, the correlation strength threshold is set as the information gain mean plus one standard deviation, and the parameter pairs below the threshold are screened out to obtain a significantly correlated process parameter group;

[0110] For the significantly correlated process parameter groups, the dynamic Bayesian network method is used to construct the parameter influence relationship diagram, and the directed connection is established through the directionality of the parameter pairs in the causal relationship table to obtain the process parameter network topology structure;

[0111] According to the topological structure of the process parameter network, conditional probability is used to represent the quantitative relationship between parameter nodes, and the connection weight coefficient is calculated by normalizing the information gain value of the parameter pair to form a weighted process parameter association network.

[0112] In the analysis of process parameter time series data, the calculation of autocorrelation function reflects the change law of parameters over time. Taking the temperature parameter of the production line as an example, a 30-minute sliding time window is used to calculate the autocorrelation coefficient of the temperature value under different time delays. When the time delay is 0, the autocorrelation coefficient is 1. As the time delay increases, the autocorrelation coefficient gradually decreases. When the delay reaches 15 minutes, the autocorrelation coefficient first appears to be negative -0.12, indicating that the temperature parameter has a 15-minute change cycle characteristic.

[0113] When determining the causal relationship between parameters, the Granger causality test method conducts a significance test by establishing a regression equation with lag terms. Taking the pressure parameter and flow parameter as an example, a regression equation containing 6 lag terms is constructed, and the calculated p value is 0.032, which is less than the significance level of 0.05, indicating that the pressure parameter is the Granger cause of the flow parameter, indicating that changes in pressure will cause changes in flow;

[0114] The conditional mutual information entropy reflects the information correlation between parameters, and the calculation formula is:

[0115] I(X;Y|Z)=H(X,Z)+H(Y,Z)-H(X,Y,Z)-H(Z)

[0116] Among them, H represents information entropy, X and Y are the parameter pairs to be tested, and Z is the conditional parameter;

[0117] Taking the temperature and pressure parameter pair as an example, under the given flow parameter conditions, the calculated conditional mutual information entropy value is 0.85, indicating that temperature and pressure have a strong correlation;

[0118] The setting of the association strength threshold is based on statistical characteristics. The information gain values ​​of all parameter pairs are calculated, and the mean is 0.6 and the standard deviation is 0.2. The information gain threshold is set to 0.8. Parameter pairs above the information gain threshold are considered to have significant correlation.

[0119] Through threshold screening, the original 15 parameter pairs were screened into 8 significantly correlated parameter pairs. In the construction of the dynamic Bayesian network, the connection weights between parameter nodes were normalized information gain values, and the calculation formula was: w = (I-min) / (max-min)

[0120] Where w is the connection weight between parameter nodes; I is the information gain value of the parameter pair; min and max are the minimum and maximum values ​​of all parameter pairs respectively;

[0121] Taking the temperature and pressure parameter pair as an example, the information gain value is 0.85, the minimum value is 0.6, and the maximum value is 0.9. The connection weight is calculated to be 0.83, indicating that temperature has a strong influence on pressure.

[0122] The parameter association analysis based on conditional probability reflects the quantitative dependence between parameters. By statistically analyzing the conditional probability distribution under different value intervals, the probability transfer matrix between parameter nodes is constructed.

[0123] Taking the effect of pressure on flow as an example, when the pressure value is in the medium range, the probability of the flow value being in the high range is 0.72, the probability of being in the medium range is 0.25, and the probability of being in the low range is 0.03, which quantitatively describes the degree of influence of pressure changes on flow distribution.

[0124] Furthermore, in this embodiment, step S3 specifically includes:

[0125] According to the real-time process parameters in the real-time process parameter data, the autocorrelation coefficient of each parameter in the real-time process parameter sequence is calculated by using a partial autocorrelation function;

[0126] The autocorrelation coefficient threshold is set according to the mean and standard deviation of the autocorrelation coefficient, and the real-time process parameter sequence exceeding the autocorrelation coefficient threshold is subjected to differential operation to obtain a stabilized process parameter sequence;

[0127] The cross-correlation coefficient between the corresponding process parameters of the new equipment and the original production equipment is calculated according to the stabilized process parameter sequence. If the cross-correlation coefficient is lower than the preset cross-correlation coefficient judgment threshold, a difference process parameter pair is obtained;

[0128] The difference process parameter pairs are decomposed by wavelet transform, and the frequency domain eigenvalues ​​are obtained by calculating the wavelet coefficients at different scales;

[0129] According to the frequency domain eigenvalues, a singular spectrum analysis method is used to select a frequency component combination that meets the preset contribution rate;

[0130] According to the combination of frequency components, a long short-term memory network structure is constructed, and the features of the real-time process parameter sequence are extracted through a sliding time window. The process parameter prediction model is obtained by training based on the features and prediction targets.

[0131] Here is an example:

[0132] The partial autocorrelation function is used to calculate the autocorrelation characteristics of each parameter in the real-time process parameter sequence. By calculating the mean and standard deviation of the autocorrelation coefficient, the autocorrelation coefficient threshold is set to the mean plus or minus two times the standard deviation. The real-time process parameter sequence that exceeds the autocorrelation coefficient threshold range is processed by first-order difference operation to obtain a stabilized process parameter sequence.

[0133] According to the stable process parameter sequence, the correlation coefficient between the corresponding process parameters of the new and old equipment is calculated by using the cross-correlation function. By setting the correlation judgment threshold to 0.7, the process parameter pairs with correlation coefficients lower than the threshold are marked to obtain a list of difference process parameter pairs.

[0134] According to the parameter sequence in the list of differential process parameters, the wavelet transform method is used to perform time-frequency decomposition, and the multi-scale frequency domain eigenvalues ​​of the process parameters are obtained by calculating the wavelet coefficients at different scales;

[0135] According to the multi-scale frequency domain eigenvalues ​​of the process parameters, the singular spectrum analysis method is used to extract the main frequency components, and the frequency component combination that meets the contribution rate is selected by setting the cumulative contribution rate threshold to 0.85;

[0136] According to the selected frequency component combination, a long short-term memory network structure with two hidden layers is constructed, the number of input layer nodes is set to be the number of frequency components, and the number of hidden layer nodes is twice and 1.5 times of the input layer respectively;

[0137] A sliding time window of size 24 is used to segment the process parameter sequence, and the parameter values ​​of the first 20 time points are used as input features, and the parameter values ​​of the last 4 time points are used as prediction targets;

[0138] The long short-term memory network was trained by back propagation algorithm, and the root mean square error was used as the loss function. When the error change rate of 5 consecutive training cycles was less than 0.001, the training was stopped to obtain the process parameter prediction model.

[0139] In the stationarity analysis of the process parameter sequence, the partial autocorrelation function reflects the sequence correlation after removing the influence of the intermediate terms. Taking the temperature parameter sequence as an example, the partial autocorrelation coefficients from lag1 to lag5 are calculated to be 0.82, 0.65, 0.43, 0.28, and 0.15, respectively, with a mean of 0.47 and a standard deviation of 0.24. The threshold interval is set to -0.01 to 0.95. Since the lag1 coefficient of the temperature sequence of 0.82 is within the threshold interval, no differential processing is required;

[0140] The cross-correlation analysis reflects the correlation between the parameters of the new and old equipment. Taking the pressure parameter as an example, the cross-correlation coefficient of the pressure series of the new and old equipment is 0.62, which is lower than the set threshold of 0.7, indicating that the pressure parameters of the new equipment are significantly different from those of the original production equipment, and subsequent analysis is needed;

[0141] At the same time, the cross-correlation coefficient of the flow parameters was calculated to be 0.85, which is higher than the threshold, indicating that the flow parameters have good consistency;

[0142] Wavelet transform realizes multi-scale decomposition of parameter sequence. Taking the pressure sequence in the difference parameter pair as an example, db4 wavelet basis function is used for three-layer decomposition to obtain three detail coefficient sequences d1, d2, d3 and one approximate coefficient sequence a3, where d1 reflects high-frequency fluctuations, accounting for 15% of the energy, d2 reflects medium-frequency fluctuations, accounting for 25% of the energy, d3 reflects low-frequency fluctuations, accounting for 35% of the energy, and a3 reflects the overall trend, accounting for 25% of the energy.

[0143] Singular spectrum analysis extracted the main frequency characteristics of the sequence, decomposed the wavelet coefficients of the pressure parameters, set the embedding dimension to 24, and calculated the singular value sequence {186.5, 142.8, 98.3, 65.7, 42.1, 28.4}, with cumulative contribution rates of {0.33, 0.58, 0.75, 0.87, 0.94, 0.99}, respectively. The first four frequency components that met the cumulative contribution rate threshold of 0.85 were selected;

[0144] In the construction of the long short-term memory network, the number of input layer nodes is determined to be 4, the number of first hidden layer nodes is 8, the number of second hidden layer nodes is 6, and the number of output layer nodes is 1 according to the 4 frequency components;

[0145] Using a 24-hour sliding time window, each sample contains 20 historical pressure values ​​as input features to predict the pressure values ​​at the next four time points;

[0146] The root mean square error was used as the loss function during the training process. The initial error was 0.245. After 123 training cycles, the error change rate for 5 consecutive cycles dropped to 0.0008, which was lower than the set threshold of 0.001, and the model training was completed.

[0147] In the prediction result verification, by comparing the model prediction value with the actual pressure value, the root mean square error of the prediction was calculated to be 0.086, and the average absolute percentage error was 3.2%, indicating that the prediction model has good prediction accuracy;

[0148] The model prediction results show that the predicted values ​​of the pressure parameters of the new equipment at the four future time points are 486.5, 492.8, 488.3, ​​and 485.7 kPa respectively;

[0149] Furthermore, the process parameter data of the new equipment including temperature, pressure, flow rate and speed are collected in real time, and the partial correlation and cross-correlation analysis of the same parameters of the new and old equipment are carried out using mathematical statistics methods, the correlation coefficient matrix is ​​calculated to determine the difference in parameter correlation, and the signal processing method is used to perform spectrum analysis on parameter fluctuations, and the key parameter combination after the introduction of the new equipment is identified through power spectrum density and coherence spectrum;

[0150] Specifically, the process parameters of temperature, pressure, flow rate and speed collected by the data collector are collected, and the missing values ​​of the process parameters are filled by the forward linear interpolation method to obtain the filled process parameters;

[0151] According to the filled process parameters, the statistical characteristic values ​​of the process parameters are calculated using a sliding time window. The statistical characteristic values ​​include mean, variance, skewness and kurtosis.

[0152] The partial correlation coefficients of similar process parameters of new and old equipment are calculated based on the statistical characteristic values ​​of process parameters, and a list of different process parameters is obtained by comparing the partial correlation coefficients with the preset thresholds.

[0153] According to the difference process parameter list, the process parameters are decomposed by wavelet transform, and the key process parameter combination is obtained by calculating the coherence coefficient of the process parameters on each frequency component;

[0154] Here is an example:

[0155] A data collector is used to obtain four types of process parameters, namely temperature, pressure, flow rate and speed, from the new equipment in real time. The process parameters are collected by setting the sampling time interval to 100 milliseconds, and the missing values ​​in the collected process parameters are filled by the forward linear interpolation method.

[0156] According to the filled process parameters, the sliding time window is used to segment the process parameters, and the statistical characteristic values ​​of the process parameters are calculated by setting the window length to 1 minute, including four statistics: mean, variance, skewness and kurtosis;

[0157] For the statistical characteristic values ​​of process parameters, the partial correlation function is used to calculate the correlation of similar process parameters of new and old equipment after removing the influence of other parameters. By calculating the mean and standard deviation of the partial correlation coefficient, the threshold of the partial correlation coefficient is set as the mean minus two times the standard deviation.

[0158] According to the partial correlation coefficient threshold, the process parameters of the new and old equipment are screened, and the process parameter pairs below the threshold are marked as difference process parameters to establish a difference process parameter list;

[0159] For the difference process parameter list, wavelet transform is used to perform three-layer time-frequency decomposition on the process parameters. By calculating the energy distribution ratio of wavelet coefficients in high frequency band, medium frequency band and low frequency band, the process parameter fluctuation characteristic sequence is obtained.

[0160] According to the characteristic sequence of process parameter fluctuations, the power spectrum density function is used to calculate the energy distribution of process parameters in the frequency domain, and the main frequency components are screened by the cumulative energy contribution rate to obtain the frequency domain characteristic vector;

[0161] For the frequency domain feature vector, the coherence spectrum function is used to calculate the coherence coefficient of the process parameters of the new and old equipment on each frequency component. By setting the coherence threshold to 0.8, the highly correlated frequency components are screened and the key process parameter combination is identified.

[0162] In process parameter collection, the selection of sampling time interval directly affects data quality. Taking temperature parameter collection as an example, a sampling interval of 100 milliseconds is used to obtain temperature data, and 600 sampling points can be obtained per minute. When a sampling point is missing, it is filled by forward linear interpolation. For example, if the 350th sampling point is missing, the temperature value of the 349th point is 185.6 degrees Celsius, and the temperature value of the 351st point is 186.2 degrees Celsius, then the interpolation calculation result of the 350th point is 185.9 degrees Celsius;

[0163] For the statistical characteristic calculation of process parameters, a 1-minute sliding time window is used for segmented statistics. Taking the pressure parameter as an example, the mean pressure in a certain time window is 500 kPa, the variance is 25, the skewness is 0.15, and the kurtosis is 3.2. These statistics comprehensively reflect the distribution characteristics and fluctuation state of the pressure parameter;

[0164] As the time window slides, the statistical characteristic values ​​form a new time series. In the partial correlation analysis, by removing the indirect influence of other parameters, the partial correlation coefficient sequence of the pressure parameters of the new and old equipment is calculated {0.92, 0.78, 0.65, 0.52, 0.43}, with a mean of 0.66 and a standard deviation of 0.18. The threshold is set to 0.30, and the parameter pairs below the threshold are marked as difference parameters;

[0165] The partial correlation coefficient sequence of the comparison flow parameters is {0.95, 0.88, 0.82, 0.75, 0.70}, with a mean of 0.82, which is higher than the threshold, indicating that the flow parameters have good consistency;

[0166] Wavelet transform realizes multi-scale analysis of parameter sequences. Taking the pressure sequence in the difference parameters as an example, after three-layer wavelet decomposition, it is found that the high-frequency band energy accounts for 25%, the medium-frequency band energy accounts for 35%, and the low-frequency band energy accounts for 40%, reflecting that the pressure fluctuation is mainly composed of medium and low frequency components.

[0167] This time-frequency decomposition helps to identify the main frequency components of parameter fluctuations. In the power spectrum density analysis, the energy distribution of the pressure parameters at different frequency points is calculated. The energy values ​​corresponding to the frequency components {0.1, 0.2, 0.3, 0.4, 0.5} Hz are {0.45, 0.25, 0.15, 0.10, 0.05}, and the cumulative energy contribution rate is {0.45, 0.70, 0.85, 0.95, 1.00}. The first three frequency components are selected as the main features;

[0168] Coherence spectrum analysis reflects the correlation degree of parameters in the frequency domain. The coherence coefficients of the pressure parameters of the new and old equipment on the main frequency components are calculated to be {0.92, 0.85, 0.75} respectively. The coherence of the first two frequency components is higher than the 0.8 threshold, indicating that these two frequency components are important bases for forming key parameter combinations. This frequency domain-based analysis method can effectively identify the main characteristic components in parameter fluctuations.

[0169] Furthermore, in this embodiment, step S4 specifically includes:

[0170] An exponential decay function is used to calculate the time weight value of the historical process parameter data, and the historical process parameter data is weighted according to the time weight value to obtain a weighted historical data set;

[0171] Divide the sliding time window according to the weighted historical data set, and obtain the data distribution feature vector by calculating the statistics of the data samples in the sliding time window;

[0172] Calculate the difference value of new and old data samples according to the data distribution feature vector, and mark the new and old data samples whose difference value exceeds the preset difference threshold as abnormal samples;

[0173] Calculate the residual size of the abnormal sample, determine the adaptive learning rate value according to the residual size, update the parameters of the process parameter prediction model through the adaptive learning rate value, and obtain the prediction result;

[0174] Here is an example:

[0175] According to the error between the parameter prediction value output by the prediction model and the actual collected value, an exponential decay function is used to Calculate the time weight of historical data by setting the time decay factor The weight value of historical data beyond the prediction interval is reduced to 0.1 to obtain a weighted historical data set;

[0176] For the weighted historical data set, a sliding time window with a length of 24 hours is used to divide the boundary between new and old data. The data distribution feature vector is obtained by calculating the mean, variance, skewness and kurtosis of the data samples in the window.

[0177] According to the data distribution feature vector, the distribution difference between the new and old data samples is calculated, and the degree of change of the data distribution is determined by setting a difference threshold. The new and old data samples whose difference values ​​exceed the preset difference threshold are marked as abnormal samples;

[0178] According to the marked abnormal samples, an adaptive learning rate calculation method based on residual is adopted to dynamically adjust the learning rate value according to the residual size, and the learning rate adjustment range is limited to between 0.001 and 0.1;

[0179] According to the adaptive learning rate value, the stochastic gradient descent method is used to update the prediction model parameters online, and the prediction model weights and bias items are iteratively adjusted by calculating the parameter gradient direction and step size;

[0180] For the adjusted forecast model parameters, the exponential smoothing method is used to calculate the forecast residual sequence, and the forecast results are weighted averaged by setting the smoothing coefficient to obtain the smoothed forecast results;

[0181] According to the smoothed prediction results, the sliding prediction error calculation method is used to evaluate the prediction accuracy in real time by calculating the root mean square error between the predicted value and the actual value, and update the prediction results of the prediction model;

[0182] In the process of process parameter prediction, the exponential decay function reflects the timeliness of historical data. Taking temperature parameter prediction as an example, when the time decay factor When it is set to 0.1, the weight of data 1 hour from the current time is 0.905, the weight of data 2 hours from the current time is 0.819, and the weight of data 3 hours from the current time is 0.741, which reflects the decreasing importance of data over time;

[0183] A 24-hour sliding time window was used for data distribution analysis. The temperature parameters in the window were calculated to have a mean of 185 degrees Celsius, a variance of 25, a skewness of 0.15, and a kurtosis of 3.2, forming a distribution feature vector {185, 25, 0.15, 3.2};

[0184] By calculating the Euclidean distance between the new data window and the old data window, the distribution difference is 0.28. If the difference threshold is set to 0.25, the current data sample is marked as an abnormal sample;

[0185] The calculation of the residual adaptive learning rate reflects the dynamic characteristics of the model update. When the prediction residual is 15 degrees Celsius, the learning rate is calculated to be 0.05 according to the residual ratio. When the residual decreases to 5 degrees Celsius, the learning rate is reduced to 0.02 accordingly, and remains within the preset range of 0.001 to 0.1. This adaptive mechanism makes the model parameter update more flexible;

[0186] In the process of stochastic gradient descent, taking the temperature prediction model as an example, the weight parameters at a certain moment are {0.82, 0.65, 0.43}, the bias term is 0.15, and the parameter gradient direction is calculated to be {-0.05, -0.03, -0.02}. Combined with the learning rate of 0.02, the updated weight parameters are {0.81, 0.64, 0.42}, and the bias term is 0.14;

[0187] The exponential smoothing method performs weighted averaging on the forecast results. The smoothing coefficient is set to 0.3. The temperature forecast value at the current moment is 188 degrees Celsius, and the smoothed value at the previous moment is 185 degrees Celsius. The smoothed forecast value at the current moment is calculated to be 185.9 degrees Celsius. The smoothing process reduces the volatility of the forecast results.

[0188] In the prediction accuracy evaluation, by calculating the root mean square error between the predicted value and the actual value at the last 10 time points, the error sequence {2.5, 2.8, 2.3, 2.7, 2.4, 2.6, 2.9, 2.5, 2.4, 2.7} was obtained, and the average error was 2.58 degrees Celsius;

[0189] This sliding error calculation method reflects the prediction performance of the model in real time. The real-time update mechanism of the prediction results forms a complete feedback loop. The prediction error realizes the dynamic adjustment of the prediction model by affecting the data weight, learning rate and model parameters, so that the model can adapt to the changes in the distribution of process parameters in a timely manner.

[0190] Judging from the temperature prediction results, the model’s 24-hour temperature prediction accuracy reached 96.5%, demonstrating good prediction performance.

[0191] Furthermore, in this embodiment, step S5 specifically includes:

[0192] A deep autoencoder is used to reduce the dimension of the real-time process parameters. By setting the number of nodes in the encoding layer to a preset range of the original parameter dimension, a compressed feature representation of the real-time process parameters is obtained.

[0193] Calculate the cosine similarity between feature vectors for the compressed feature representation, merge the feature vectors whose cosine similarity is higher than a preset similarity threshold, and obtain a fused feature vector set;

[0194] Perform recursive feature elimination operation on the fused feature vector set, remove features whose feature importance scores are lower than the importance threshold, and obtain key feature combinations;

[0195] Generate a sampling point sequence within the historical value range of process parameters based on the key feature combination, and obtain the optimal process parameter configuration set suitable for the new equipment by performing Monte Carlo simulation calculation and multi-objective evaluation function operation;

[0196] Here is an example:

[0197] According to the incrementally updated process parameter prediction model, a five-layer deep autoencoder is used to extract the process parameter features of the new equipment. By setting the number of encoding layer nodes to three-quarters, one-half, and one-quarter of the original parameter dimension, the process parameters are encoded with reduced dimensions to obtain the compressed feature representation of the process parameters of the new equipment.

[0198] According to the compressed feature representation, the cosine similarity is used to calculate the similarity between feature vectors. By setting the similarity threshold to 0.85, the feature vectors above the threshold are merged to obtain a fused feature vector set.

[0199] According to the fused feature vector set, the recursive feature elimination method is used to calculate the feature importance score. By setting the feature importance threshold as the average score plus one standard deviation, the features below the threshold are eliminated to obtain the key feature combination.

[0200] According to the key feature combination, the Latin hypercube sampling method is used to generate sampling points within the historical value range of the process parameters. By evenly dividing the value range of each parameter, a set of parameter configuration solutions is constructed.

[0201] According to the parameter configuration scheme set, the Monte Carlo method is used to generate simulation data. By setting the single simulation time to ten times the production cycle, simulation calculation is performed on each set of parameter configurations.

[0202] According to the simulation calculation results, a multi-objective evaluation function is used to comprehensively score the product quality index, production efficiency index and resource consumption index, and the weighted score is calculated by setting the index weights;

[0203] According to the weighted score sequence, the parameter configuration schemes are screened by using the ranking selection method, and the optimal process parameter configuration set is obtained by retaining the configuration schemes with scores in the top ten percent;

[0204] In the process parameter feature extraction, a five-layer deep autoencoder is used to achieve parameter dimension reduction. Taking the four process parameters of temperature, pressure, flow rate and speed as examples, the original input dimension is 4-dimensional, the number of nodes in the encoding layer is 3, 2, and 1 respectively, and the number of nodes in the decoding layer is 2, 3, and 4 respectively. Through training, the 1-dimensional feature representation of the middle layer is obtained, which realizes the compression of the parameter dimension.

[0205] In the process of feature vector fusion, the cosine similarity calculation reflects the correlation between features. Taking two feature vectors v1={0.8, 0.5, 0.3} and v2={0.75, 0.45, 0.35} as examples, the cosine similarity calculated is 0.89, which is higher than the set threshold of 0.85. The weighted average of these two feature vectors is used to obtain the fused feature vector {0.78, 0.48, 0.32}.

[0206] In the feature importance evaluation, the recursive feature elimination method iteratively calculates the feature scores and obtains the importance score sequence {0.85, 0.72, 0.45, 0.38} for the four process parameters, with a mean of 0.60, a standard deviation of 0.20, and a threshold of 0.80. The two important feature parameters of temperature and pressure are retained. Latin hypercube sampling achieves uniform coverage of the parameter space. Taking the temperature parameter as an example, the range of 180 to 200 degrees Celsius is evenly divided into 10 sub-intervals, and 1 point is randomly sampled in each sub-interval to obtain 10 sampling values ​​of the temperature parameter {182.5, 185.8, 188.3, ​​191.2, 193.6, 194.8, 196.5, 197.9, 198.8, 199.5};

[0207] During the simulation calculation process, multiple sets of simulation data were generated through the Monte Carlo method. The single simulation duration was set to 240 minutes, covering 10 times the full production cycle of 24 minutes. Each set of parameter configurations was simulated 100 times to obtain stable simulation results.

[0208] The multi-objective evaluation adopts the weighted summation method, setting the product quality index weight to 0.4, the production efficiency index weight to 0.35, and the resource consumption index weight to 0.25. The three index values ​​of a certain set of parameter configurations {0.92, 0.88, 0.85} are weighted and calculated to obtain a comprehensive score of 0.888;

[0209] In the parameter configuration scheme screening, the comprehensive scores of 100 configuration schemes were arranged in descending order, with scores ranging from 0.75 to 0.92. The top 10% of the configuration schemes were selected, with a corresponding score threshold of 0.88, and 10 preferred configuration schemes were screened out;

[0210] The temperature parameters of these schemes are mainly distributed in the range of 190 to 195 degrees Celsius, and the pressure parameters are mainly distributed in the range of 450 to 500 kPa, which reflects the optimal value range of process parameters;

[0211] Furthermore, the real-time process parameters of the new equipment during operation, including temperature, pressure, liquid level and composition, are obtained through industrial Ethernet. The Pearson correlation coefficient is used to measure the partial correlation and cross-correlation between the parameters of the new equipment and the original association network model. The wavelet analysis method is used to perform spectral analysis on the dynamic characteristics of the parameters. The key parameter combinations that need to be optimized and adjusted under the influence of the new equipment are determined through characteristic frequency extraction and cluster analysis.

[0212] Specifically, the temperature, pressure, liquid level and composition process parameters are collected through the industrial Ethernet, the process parameter sequence is obtained by setting the sampling interval, and the data integrity check is performed on the process parameter sequence;

[0213] According to the process parameter sequence after the integrity check, the median filtering method is used to obtain the filtered process parameters, and the abnormal values ​​are judged by calculating the mean and standard deviation of the filtered process parameters;

[0214] According to the process parameters after removing the outliers, the Pearson correlation coefficient is used to calculate the correlation value, and the correlation coefficient threshold is determined by calculating the mean and standard deviation of the correlation coefficient;

[0215] According to the process parameters that are higher than the correlation coefficient threshold, wavelet decomposition is used to obtain the decomposition coefficients of the process parameters, and the frequency domain characteristics are determined by calculating the power spectrum density function of the decomposition coefficients. The parameters with similar frequency domain characteristics are merged using the hierarchical clustering method.

[0216] Here is an example:

[0217] Industrial Ethernet is used to collect four types of process parameters, namely temperature, pressure, liquid level and composition, during the operation of the new equipment. The data is collected by setting the sampling interval to 100 milliseconds, and the data integrity of the collected process parameters is checked.

[0218] The process parameters after integrity check are denoised using the median filtering method. By calculating the mean and standard deviation of the parameter sequence, the values ​​that exceed the range of three times the standard deviation are marked as outliers and filtered out.

[0219] According to the process parameters after filtering out abnormal values, the partial correlation and cross-correlation values ​​of the process parameters of the new and old equipment are calculated using the Pearson correlation coefficient. By calculating the mean and standard deviation of the correlation coefficient, the correlation coefficient threshold is set to the mean plus one standard deviation.

[0220] For the process parameter pairs that are higher than the correlation coefficient threshold, the db4 wavelet basis function is used to perform three-layer wavelet decomposition, and the decomposition coefficients of the process parameters at different frequency scales are obtained by calculating the approximate coefficients and detail coefficients.

[0221] According to the wavelet decomposition coefficient, the power spectrum density function is used to calculate the energy distribution characteristics of the process parameters in the frequency domain. By setting the frequency component energy contribution rate threshold to 0.85, the main frequency components that meet the energy contribution rate are selected.

[0222] According to the main frequency components, the hierarchical clustering method is used to calculate the spectral characteristic distance between process parameters. By setting the clustering distance threshold to 0.3 times the height of the clustering tree, the parameters with a distance less than the threshold are merged;

[0223] According to the merged process parameter group, the mutual information entropy method is used to calculate the information correlation between the parameters. By setting the information correlation threshold to 0.7, the key process parameter combination is screened;

[0224] In industrial Ethernet data collection, the setting of sampling interval directly affects the data quality. Taking temperature parameter collection as an example, a sampling interval of 100 milliseconds is set, 10 data points are obtained per second, and 600 data points can be collected in 1 minute. The data integrity check shows that 585 data points are actually collected, with a completeness rate of 97.5%;

[0225] The outlier processing adopts the median filtering method to process the temperature parameter sequence. The original data has a mean of 185 degrees Celsius and a standard deviation of 5 degrees Celsius. The outlier judgment range is set to 170 to 200 degrees Celsius. Five outliers are found, which are 165, 168, 202, 205, and 208 degrees Celsius. These outliers are corrected to the median values ​​of the adjacent points through median filtering.

[0226] The Pearson correlation coefficient calculation reflects the degree of linear correlation between parameters. Taking temperature and pressure parameters as an example, the correlation coefficient sequence {0.85, 0.78, 0.72, 0.68, 0.65} is calculated, with a mean of 0.736 and a standard deviation of 0.074. The threshold is set to 0.81, and the temperature and pressure pair of highly correlated parameters are screened out;

[0227] Wavelet decomposition uses db4 wavelet basis function to perform three-layer decomposition. Taking pressure parameters as an example, three-layer detail coefficients d1, d2, d3 and approximate coefficient a3 are obtained, where d1 reflects high-frequency fluctuations, accounting for 15% of the energy, d2 reflects medium-frequency fluctuations, accounting for 25% of the energy, d3 reflects low-frequency fluctuations, accounting for 35% of the energy, and a3 reflects signal trends, accounting for 25% of the energy.

[0228] The power spectrum density analysis shows the energy distribution of pressure parameters in the frequency domain. The energy values ​​corresponding to the main frequency components {0.1, 0.2, 0.3, 0.4, 0.5} Hz are {0.45, 0.25, 0.15, 0.10, 0.05}, and the cumulative energy contribution rate is {0.45, 0.70, 0.85, 0.95, 1.00}. The first three frequency components are selected as the main features;

[0229] In the hierarchical cluster analysis, the Euclidean distance between the process parameters was calculated, and the distance between temperature and pressure in the distance matrix was 0.25, the distance between temperature and liquid level was 0.85, the distance between temperature and composition was 0.95, the cluster tree height was 1.2, and the threshold was set to 0.36, and the temperature and pressure parameters were combined into one group;

[0230] The mutual information entropy calculation quantifies the nonlinear correlation between parameters. The mutual information entropy is calculated for the parameter combination {temperature-pressure, liquid level-composition}, and the values ​​are 0.82 and 0.45 respectively. The threshold is set at 0.7, and finally temperature-pressure is determined as the key parameter combination.

[0231] This combination shows a strong correlation in both frequency domain characteristics and information correlation, indicating that these two parameters have a close interaction in the production process.

[0232] Furthermore, in this embodiment, step S6 specifically includes:

[0233] A proportional-integral-differential controller is used to obtain the deviation mean and standard deviation of the real-time process parameters, and the proportional coefficient value is obtained according to the inverse of the standard deviation;

[0234] Collect product quality index data according to the real-time process parameter values, and obtain hardness index values, strength index values ​​and purity index values ​​through the data collector;

[0235] The center line value and process standard deviation are calculated for the quality index data. If the quality index data is within the preset standard deviation interval of the center line value, the frequency domain characteristic spectrum is obtained by Fourier transform;

[0236] The frequency domain characteristic spectrum is decomposed by three layers of wavelet, and the quality evaluation score is obtained by calculating the energy value of the high frequency band, the energy value of the middle frequency band and the energy value of the low frequency band. The control parameters of the process parameters are dynamically adjusted according to the quality evaluation score.

[0237] Here is an example:

[0238] According to the process parameter configuration plan, a proportional-integral-differential controller is used to adjust the four process parameters of the new equipment, namely temperature, pressure, liquid level and composition, in real time. By calculating the mean and standard deviation of the parameter deviation, the proportional coefficient is set to the inverse of the standard deviation, the integral time is 4 times the sampling period, and the differential time is one quarter of the sampling period.

[0239] For the adjusted process parameters, a data collector is used to collect three types of quality index data of product hardness, strength and purity, and the quality index data are preprocessed by setting the sampling interval to 1 minute;

[0240] According to the pre-processed quality index data, the statistical process control card is used for monitoring. By calculating the center line value of the quality index and the process standard deviation, the upper and lower control limits are set as the center line value plus or minus three times the standard deviation;

[0241] According to the quality index data within the control limit interval, the frequency domain characteristic spectrum is calculated by Fourier transform, and the frequency spectrum analysis of the quality index data is performed by setting the frequency resolution to one thousandth of the sampling frequency;

[0242] According to the frequency domain characteristic spectrum, the quality index data is decomposed into three layers by wavelet transform, and the time-frequency characteristic value of the quality index is obtained by calculating the energy distribution of the three frequency bands of high frequency, medium frequency and low frequency.

[0243] According to the time-frequency eigenvalues, the weighted summation method was used to calculate the comprehensive quality evaluation score, and the quality evaluation results were obtained by setting the weight ratio of the three indicators of hardness, strength and purity to 4:3:3;

[0244] According to the quality evaluation results, the adaptive feedback compensation algorithm is used to correct the process parameter configuration plan, and the control parameters of the process parameters are dynamically adjusted by calculating the deviation between the quality score and the target value;

[0245] In the process of process parameter control, the parameter setting of the proportional-integral-differential controller directly affects the control effect. Taking the temperature parameter as an example, when the mean value of the temperature deviation is 2 degrees Celsius and the standard deviation is 0.5 degrees Celsius, the proportional coefficient is set to 2, the sampling period is 10 seconds, the integral time is set to 40 seconds, and the differential time is set to 2.5 seconds. This parameter configuration enables the controller to have appropriate response speed and stability.

[0246] During the quality indicator data collection, the product hardness was monitored, the sampling interval was set to 1 minute, 60 data points were collected continuously, and the calculated centerline value was 85HV (Vickers hardness), the process standard deviation was 2HV, and the upper limit was set to 91HV and the lower limit was 79HV. At the same time, the product strength was monitored, and the tensile strength centerline value was 450MPa, the standard deviation was 15MPa, the upper limit was 495MPa, and the lower limit was 405MPa;

[0247] In the frequency domain feature analysis, the sampling frequency was set to 1 Hz, the frequency resolution was 0.001 Hz, and the hardness data was subjected to Fourier transform, and the main frequency components were distributed in the range of 0.01 to 0.05 Hz, reflecting the periodic characteristics of hardness changes;

[0248] Analysis of the strength data revealed that the main frequency components were in the range of 0.02 to 0.08 Hz, indicating that the strength fluctuation frequency was slightly higher than that of the hardness;

[0249] Wavelet transform uses three-level decomposition. Taking the purity index as an example, the energy of the high-frequency band (0.25-0.5 Hz) accounts for 15%, reflecting short-term fluctuations;

[0250] The energy in the mid-frequency band of 0.125-0.25 Hz accounts for 35%, reflecting mid-term changes;

[0251] The energy in the low frequency band 0-0.125 Hz accounts for 50%, reflecting the long-term trend;

[0252] This time-frequency analysis reveals the multi-scale characteristics of purity fluctuations. In the comprehensive quality evaluation, the weighted summation method is used to score the hardness index of a batch of products at 88HV, the strength index at 465MPa, and the purity index at 98.5%. The weights are 0.4, 0.3, and 0.3, respectively. The calculated comprehensive score is 0.92, which is -0.03 compared with the target value of 0.95. The process parameters need to be adjusted.

[0253] Adaptive feedback compensation dynamically adjusts control parameters according to quality deviation. When the quality score is lower than the target value, the proportional coefficient of temperature control is increased to 2.2, the integration time is shortened to 35 seconds, and the response strength of the controller is appropriately improved.

[0254] At the same time, the pressure control parameters were adjusted, the proportional coefficient was increased to 1.8, the integration time was extended to 45 seconds, and the coordinated control of process parameters was strengthened. Through this adaptive adjustment mechanism, dynamic optimization of quality indicators was achieved.

[0255] Furthermore, in this embodiment, step S7 specifically includes:

[0256] The update amount of the parameter association matrix is ​​calculated based on the process parameters and quality index data. The update amount is determined by the incremental gradient descent algorithm and the baseline learning rate.

[0257] According to the parameter correlation matrix, the sliding time window method is used to obtain the probability distribution of new and old samples, and the chi-square distance value of the sample distribution is calculated;

[0258] According to the chi-square distance value, the random forest algorithm is used to build a process parameter optimizer. The process parameter optimizer obtains the optimal split feature by calculating the Gini index, and the structure of the process parameter optimizer is incrementally updated according to the optimal split feature.

[0259] According to the process parameter optimizer, a proportional-integral controller is used to calculate the parameter adjustment amount, and the process parameters are adjusted in real time according to the parameter adjustment amount;

[0260] Here is an example:

[0261] According to the process parameters and quality index data generated by the operation of the new equipment, the incremental gradient descent algorithm is used to calculate the update amount of the parameter association matrix. By setting the baseline learning rate to 0.01 and dynamically adjusting the step size based on the number of batch samples, the association matrix is ​​updated online.

[0262] For the updated correlation matrix, the sliding time window method is used to calculate the probability distribution of new and old samples, and the degree of sample distribution difference is quantitatively evaluated by calculating the chi-square distance value between the distributions;

[0263] According to the sample distribution difference value, the adaptive learning rate calculation formula is used to dynamically adjust the update step size of incremental learning by setting the upper and lower limits of the learning rate to 0.1 and 0.001 respectively;

[0264] For the adjusted learning rate, the random forest algorithm is used to build the process parameter optimizer, the optimal splitting feature is determined by calculating the Gini index, and the optimizer structure is incrementally updated;

[0265] According to the updated optimizer structure, the entropy-based feature importance evaluation method is used to rank the importance of process parameters by calculating the information gain ratio to obtain the parameter optimization order;

[0266] For the parameter optimization sequence, a proportional-integral controller is used to achieve closed-loop regulation, and the process parameters are adjusted in real time by calculating the mean and integral value of the deviation between the actual parameter value and the target value;

[0267] According to the parameter adjustment results, the prediction residual of the quality index is calculated by using the exponential weighting method, and the parameter optimization scheme is dynamically evaluated and corrected by setting the weighting coefficient to 0.9;

[0268] In the online update process of the process parameter association matrix, the incremental gradient descent algorithm calculates the update amount through batch samples. Taking the temperature and pressure parameters as an example, the initial association weight is 0.65. When the batch sample size is 50, the calculated gradient direction is -0.08, the baseline learning rate is 0.01, and the actual step size is , the updated association weight becomes 0.64;

[0269] The quantitative evaluation of the sample distribution difference uses a sliding time window method with a window length of 60 minutes to calculate the probability distribution of new and old samples;

[0270] The temperature parameter is calculated to have a new sample mean of 185 degrees Celsius and a standard deviation of 4.5 degrees Celsius, and an old sample mean of 182 degrees Celsius and a standard deviation of 3.8 degrees Celsius. The chi-square distance calculation yields a distribution difference value of 0.28;

[0271] The dynamic adjustment of the adaptive learning rate is based on the sample distribution difference value. When the difference value is greater than 0.3, the learning rate increases to 0.08;

[0272] When the difference value is less than 0.1, the learning rate is reduced to 0.003;

[0273] When the difference value is between 0.1 and 0.3, the learning rate is kept at 0.01;

[0274] This adaptive mechanism makes the model update speed match the degree of data change. During the incremental update process of the random forest optimizer, the optimal split feature is selected by calculating the Gini index. Taking the pressure parameter as an example, the Gini index of different split points is calculated {0.42, 0.38, 0.45, 0.41}, and the split point corresponding to the minimum Gini index of 0.38 is selected as the decision node;

[0275] With the addition of new data, the decision tree structure is continuously optimized. The feature importance evaluation adopts the information gain ratio calculation method. The information gain ratios of the three parameters of temperature, pressure and liquid level are calculated to be 0.45, 0.32 and 0.28 respectively. The parameter optimization order is determined to be temperature, pressure and liquid level.

[0276] This ranking reflects the degree of influence of different parameters on the production process. In the closed-loop control process, the proportional-integral controller is adjusted according to the parameter deviation. Taking temperature control as an example, when the actual temperature is 185 degrees Celsius and the target temperature is 190 degrees Celsius, the calculated deviation mean is -5 degrees Celsius and the integral value is -15 degrees Celsius per minute. Based on this, the control output value is calculated to adjust the temperature.

[0277] The predicted residual calculation of quality indicators adopts the exponential weighting method. For the product hardness indicator, the current residual is 2HV, the historical weighted residual is 1.5HV, and the weighting coefficient is 0.9. The new weighted residual is 1.55HV. The parameter adjustment strategy is continuously optimized through residual analysis to achieve dynamic optimization of process parameters.

[0278] The present invention also proposes a data processing device based on artificial intelligence technology and a data model, comprising:

[0279] Data collection and preprocessing module, used to collect historical process parameter data of original production equipment and real-time process parameter data of new equipment, and perform data preprocessing;

[0280] Model building module, used to build correlation network model and process parameter prediction model;

[0281] Model optimization and application module, which is used to integrate the real-time process parameter feature representation of new equipment into the association network model, obtain the optimal process parameter configuration set suitable for the new equipment, apply the optimal process parameter configuration set to the actual production process, and perform dynamic optimization through the feedback mechanism;

[0282] The knowledge accumulation and closed-loop control module is used to continuously update the associated network model and the optimization decision knowledge base, continuously accumulate and improve the adaptive knowledge introduced by new equipment, and form a dynamic optimization closed-loop control mechanism;

[0283] The data processing device based on artificial intelligence technology and data model includes the data processing method based on artificial intelligence technology and data model as described above.

[0284] The present invention also proposes a data processing device based on artificial intelligence technology and a data model, comprising:

[0285] at least one memory;

[0286] at least one processor;

[0287] at least one program;

[0288] The program is stored in the memory, and the processor executes at least one program to implement the data processing method based on artificial intelligence technology and data model as described above.

[0289] In the description of the present invention, it is necessary to understand that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "lateral, vertical, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific direction or be constructed and operated in a specific direction. Therefore, they cannot be understood as limiting the scope of protection of the present invention.

[0290] For those skilled in the art, various other corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all of these changes and deformations should fall within the protection scope of the claims of the present invention.

Claims

1. A data processing method based on artificial intelligence technology and data model, characterized in that: The following steps are involved: S1. Obtain the historical process parameter data of the original production equipment and the real-time process parameter data of the new equipment, and perform data preprocessing to obtain a standardized time series process parameter data set; S2. By mining the dynamic correlation characteristics between different process parameters and determining the causal relationship between the parameters, a correlation network model reflecting the intensity of mutual influence of the parameters is constructed; S3. Conduct partial correlation analysis and cross-correlation analysis on the real-time process parameters of the new equipment and the historical process parameters of the original production equipment, and build a process parameter prediction model; S4, incrementally updating the process parameter prediction model, and using an adaptive learning algorithm to adjust the parameters of the process parameter prediction model and the weight of evaluating new samples; S5. Through the knowledge increment fusion method, the real-time process parameter feature representation of the new equipment is integrated into the association network model, and the parameter combination is optimized to obtain the optimal process parameter configuration set suitable for the new equipment; S6. Applying the optimal process parameter configuration set to the actual production process, adjusting and monitoring the parameters of the new equipment in real time, and dynamically correcting the optimal process parameter configuration set through a feedback mechanism; S7. Utilize an incremental learning algorithm to continuously update the association network model and the optimization decision knowledge base, continuously accumulate and improve the adaptive knowledge introduced by the new equipment, and form a dynamic optimization closed-loop control mechanism.

2. The data processing method based on artificial intelligence technology and data model according to claim 1 is characterized in that: The step S1 specifically includes: Based on the historical process parameter data of the original production equipment and the real-time process parameter data of the new equipment, data inspection is performed by setting a time window with a fixed time interval; For process parameter data with missing values ​​in parameter records, linear interpolation method is used to obtain supplementary filling data; Calculate the mean and standard deviation of the parameter values ​​according to the supplementary filling data, and obtain the outlier elimination result according to whether the parameter exceeds the preset standard deviation interval; According to the abnormal value elimination result, a parameter type comparison table of temperature parameters, pressure parameters and flow parameters is established, and dimensionally unified parameters are obtained through Kelvin conversion, Pascal conversion and cubic meters per second conversion; The dimensionally unified parameters are normalized to obtain normalized values, and the Pearson correlation coefficient is calculated for the normalized values ​​using a sliding time window to determine the process parameter mapping relationship to obtain a standardized time series process parameter data set.

3. The data processing method based on artificial intelligence technology and data model according to claim 1 is characterized in that: The step S2 specifically includes: According to the process parameter time series data in the time series process parameter data set, a parameter autocorrelation coefficient sequence is calculated using a sliding window autocorrelation function, and the parameter change period is determined by identifying the time delay value corresponding to the first negative value of the autocorrelation coefficient; According to the parameter variation cycle, the Granger causality test method is used to calculate the significance level value of the parameter pair. If the significance level value is less than the preset significance threshold, it is determined that there is a causal relationship between the parameter pair; For the parameter pairs with causal relationship, the conditional mutual information entropy method is used to calculate the information gain value, and the dynamic correlation strength value of the parameter pair is obtained by comparing the information gain value with a preset information gain threshold; According to the dynamic association strength value, a dynamic Bayesian network method is used to construct a parameter influence relationship diagram, and a network connection weight coefficient is established through the directionality of the parameter pairs in the causal relationship table.

4. The data processing method based on artificial intelligence technology and data model according to claim 1 is characterized in that: The step S3 specifically includes: According to the real-time process parameters in the real-time process parameter data, a partial autocorrelation function is used to calculate the autocorrelation coefficient of each parameter in the real-time process parameter sequence; An autocorrelation coefficient threshold is set according to the mean and standard deviation of the autocorrelation coefficient, and a difference operation is performed on the real-time process parameter sequence exceeding the autocorrelation coefficient threshold to obtain a stabilized process parameter sequence; Calculating the cross-correlation coefficient between the corresponding process parameters of the new equipment and the original production equipment according to the stabilized process parameter sequence, and obtaining a difference process parameter pair if the cross-correlation coefficient is lower than a preset cross-correlation coefficient determination threshold; Decomposing the difference process parameter pair by wavelet transform, and obtaining frequency domain eigenvalues ​​by calculating wavelet coefficients at different scales; According to the frequency domain eigenvalues, a frequency component combination satisfying a preset contribution rate is selected using a singular spectrum analysis method; According to the frequency component combination, a long short-term memory network structure is constructed, and the features of the real-time process parameter sequence are extracted through a sliding time window. The process parameter prediction model is obtained according to the feature and prediction target training.

5. The data processing method based on artificial intelligence technology and data model according to claim 1 is characterized in that: The step S4 specifically includes: An exponential decay function is used to calculate a time weight value of the historical process parameter data, and the historical process parameter data is weighted according to the time weight value to obtain a weighted historical data set; Dividing a sliding time window according to the weighted historical data set, and obtaining a data distribution feature vector by calculating the statistics of data samples in the sliding time window; Calculating the difference value of the new and old data samples according to the data distribution feature vector, and marking the new and old data samples whose difference value exceeds a preset difference threshold as abnormal samples; The residual size of the abnormal sample is calculated, an adaptive learning rate value is determined according to the residual size, parameters of the process parameter prediction model are updated by the adaptive learning rate value, and a prediction result is obtained.

6. The data processing method based on artificial intelligence technology and data model according to claim 1 is characterized in that: The step S5 specifically includes: A deep autoencoder is used to perform dimension reduction encoding on the real-time process parameters, and a compressed feature representation of the real-time process parameters is obtained by setting the number of encoding layer nodes to a preset range of the original parameter dimension; Calculating the cosine similarity between feature vectors for the compressed feature representation, merging feature vectors whose cosine similarity is higher than a preset similarity threshold, and obtaining a fused feature vector set; Performing a recursive feature elimination operation on the fused feature vector set, eliminating features whose feature importance scores are lower than an importance threshold, and obtaining a key feature combination; A sampling point sequence is generated within the historical value interval of the process parameters according to the key feature combination, and an optimal process parameter configuration set suitable for the new equipment is obtained by executing Monte Carlo simulation calculation and multi-objective evaluation function operation.

7. The data processing method based on artificial intelligence technology and data model according to claim 1 is characterized in that: The step S6 specifically includes: A proportional-integral-differential controller is used to obtain a deviation mean and a standard deviation of the real-time process parameter, and a proportional coefficient value is obtained according to the inverse of the standard deviation; Collect product quality index data according to the numerical value of the real-time process parameter, and obtain hardness index value, strength index value and purity index value through a data collector; Calculating the center line value and the process standard deviation of the quality indicator data, and if the quality indicator data is within the preset standard deviation interval of the center line value, using Fourier transform to obtain the frequency domain characteristic spectrum; The frequency domain characteristic spectrum is subjected to three-layer wavelet decomposition, and a quality evaluation score is obtained by calculating the high-frequency band energy value, the mid-frequency band energy value and the low-frequency band energy value, and the control parameters of the process parameters are dynamically adjusted according to the quality evaluation score.

8. The data processing method based on artificial intelligence technology and data model according to claim 1 is characterized in that: The step S7 specifically includes: Calculating an update amount of a parameter association matrix according to process parameters and quality indicator data, wherein the update amount is determined by an incremental gradient descent algorithm and a baseline learning rate; According to the parameter association matrix, a sliding time window method is used to obtain the probability distribution of new and old samples, and the chi-square distance value of the sample distribution is calculated; According to the chi-square distance value, a process parameter optimizer is constructed by using a random forest algorithm, the process parameter optimizer obtains an optimal split feature by calculating a Gini index, and the structure of the process parameter optimizer is incrementally updated according to the optimal split feature; According to the process parameter optimizer, a proportional-integral controller is used to calculate the parameter adjustment amount, and the process parameters are adjusted in real time according to the parameter adjustment amount.

9. A data processing device based on artificial intelligence technology and data model, characterized in that: include: Data collection and preprocessing module, used to collect historical process parameter data of original production equipment and real-time process parameter data of new equipment, and perform data preprocessing; Model building module, used to build correlation network model and process parameter prediction model; The model optimization and application module is used to integrate the real-time process parameter feature representation of the new equipment into the association network model, obtain the optimal process parameter configuration set suitable for the new equipment, apply the optimal process parameter configuration set to the actual production process, and perform dynamic optimization through the feedback mechanism; The knowledge accumulation and closed-loop control module is used to continuously update the associated network model and the optimization decision-making knowledge base, continuously accumulate and improve the adaptive knowledge introduced by new equipment, and form a dynamically optimized closed-loop control mechanism.

10. Data processing equipment based on artificial intelligence technology and data model, characterized in that: include: at least one memory; at least one processor; at least one program; The program is stored in the memory, and the processor executes at least one of the programs to implement the data processing method based on artificial intelligence technology and data model as described in any one of claims 1 to 8.

Images