Rapid and accurate arc current control method based on plasma load current situation awareness
By predicting and controlling the load current changes during the sputtering process based on plasma load current situation awareness, the problem of arcing is solved, and the stability of the coating process and the film layer quality are improved.
Patent Information
- Application Number
- CN202411065194.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-05
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
During magnetron sputtering vacuum coating, load current disturbance causes arcing, affecting film uniformity and may lead to production interruption or power supply damage.
The fast and precise arc current control method based on plasma load current situation perception is adopted. By collecting real-time current data, voltage data, frequency data, pressure in the cavity, gas flow rate and ambient temperature, the time series is constructed, and the CNN-GRU model algorithm is used. Predict the load current change trend and fuse the predicted value with traditional proportional integral control to achieve suppression of current spikes.
Effectively predict the change trend of load current, regulate the current in advance, eliminate arcs, avoid power protection status, ensure constant load current during the coating process, and improve the quality of the deposited film layer.
Smart Images

Figure CN118996366A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of magnetron sputtering vacuum coating, and in particular to a method for fast and accurate arc current control based on plasma load current situation perception. Background Art
[0002] As a deposition coating method, magnetron sputtering technology is widely used in industrial production and scientific research fields, such as microelectronics, optical thin films and material surface treatment. With the continuous development of science and technology, people have put forward higher requirements on the wear resistance, corrosion resistance, high temperature resistance and oxidation resistance of the materials used. The thin film produced by sputtering coating technology has the advantages of strong bonding, dense film layer, uniformity, no pollution, low cost and high efficiency, and has received more and more attention and research in countries around the world.
[0003] Advanced arc management system is the basis of sputtering process and the guarantee of high sputtering rate. During the sputtering process, due to the large-scale continuous changes in load and environment, high requirements are placed on the output performance of the power supply. Whether the stability and reliability of the power supply device can adapt to the above changes and provide good sputtering conditions seriously affects the quality of the film. In particular, due to the continuous accumulation of charge on the cathode during the sputtering process, an arc is inevitably formed between the target and the anode, and arcing occurs. If the formed arc cannot be suppressed in time, it is easy to cause film defects, affect the uniformity of the film, and appear spots and pinholes on the surface. In severe cases, production will be forced to be interrupted or even the power supply will be damaged. Therefore, in order to ensure the quality of the coating, the power supply must have a good "stable current" capability, quickly and accurately control the arc current and suppress disturbances, so as to ensure the stability of the sputtering deposition process. Summary of the invention
[0004] 1. Technical issues to be resolved
[0005] In view of the deficiencies in the prior art, the present invention provides a method for fast and accurate arc current control based on plasma load current situation awareness, which solves the problem of arc generation caused by load current disturbance during sputtering deposition.
[0006] (II) Technical solution
[0007] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a method for fast and accurate arc current control based on plasma load current situation awareness, comprising the following steps:
[0008] Step 1: Load current factor extraction
[0009] In the vacuum coating deposition operation environment, the load operation status information is collected, including real-time current data, voltage data, frequency data, cavity pressure, gas flow and ambient temperature information. The time series is constructed based on this information, which is the data basis for realizing current situation awareness.
[0010] Step 2: Understanding the load current situation
[0011] Deeply explore the implicit correlation and data specificity in the data information provided by the load current element extraction, and fully understand the load current variation law, which is a deep understanding of the extracted information;
[0012] Step 3: Load current situation prediction
[0013] The CNN-GRU model algorithm is used to learn the high-level features of the data after situation understanding, to mine the before and after relationships of long time series, and to predict the trend of load current changes;
[0014] Step 4: Fast and accurate arc current control strategy
[0015] The predicted value is integrated with the traditional proportional-integral control to form a composite control with a feedforward function. The current spike is suppressed through load current feedforward to prevent the power supply from entering the protection state, thereby ensuring the constant load current in the process and improving the quality of the deposited film.
[0016] Preferably, the switch power supply of the vacuum coating in step 1 is equipped with a sensor unit and a data communication module with the coating equipment to construct the time series of three electrical signals of voltage U, current I and frequency F at time t, and the time series of key parameters of the coating equipment such as pressure P in the cavity, gas flow Q and ambient temperature K:
[0017] V t =[v1,v2,…,v j ,…] Τ ,j∈Ν (1)
[0018] I t =[i1,i2,…,i j ,…] Τ ,j∈Ν (2)
[0019] F t =[f1,f2,…,f j ,…] Τ ,j∈Ν (3)
[0020] P t =[p1,p2,…,p j ,…] Τ ,j∈Ν (4)
[0021] Q t =[q1,q2,…,q j ,…] Τ ,j∈Ν (5)
[0022] K t=[k1,k2,…,k j ,…] Τ ,j∈Ν (6)
[0023] Since the sampling time of the sensor electrical signal is not synchronized with the data time of the key parameters of the coating equipment, the time scales of equations (1) to (6) are inconsistent. In order to avoid the above situation, the average value of the collected data is extracted at a certain interval time as a reference value according to the situation prediction time scale. Taking the time period current I and pressure P sequence as an example, the calculation method is shown in equations (7) to (8):
[0024]
[0025] In formula (7) and formula (8), i j1 and p j1 Corresponding to time period t j-1 +Δt current and pressure, i jm and p jm Corresponding to time period t j +Δt~t j +mΔt, and on this basis, the average current and pressure at time j are obtained, thereby deriving the current and pressure input time series of the model.
[0026] Preferably, in step 2, outliers need to be processed. First, an unsupervised learning method is used to detect outliers, including data density and cluster analysis, that is, normal values and outliers are distinguished based on the characteristics of the time series data itself, using the load current sequence data density and the distance between the load current sequence data; secondly, for the processing of abnormal collected values, the continuity of the load current data and the external factor data in the time series is ensured, and the missing values are filled by the local weighted regression interpolation method, while the abnormal deviation values are first removed and then locally weighted regression interpolation is performed.
[0027] Preferably, in step 2, the load current data and the external factor data are subjected to value standardization processing, and the calculation is as follows:
[0028]
[0029] In the formula, q t ' represents the standardized data of the value at time t, q t represents the original data at time t, q max and q min Respectively represent the maximum and minimum values in the time series data of each dimension.
[0030] Preferably, the load current data in step 2 is used as a time series data, and its correlation in time series needs to be analyzed first to provide the input time step of the situation prediction model to obtain the best prediction effect, that is, the historical data within a certain number of seconds from the current moment is used as input. The autocorrelation coefficient C(l) can measure the correlation degree of the load current time series data at two different time points. The value range of C(l) is [-1, 1]. If C(l)>0, the two load current time series data are positively correlated, otherwise they are negatively correlated; the larger |C(l)| is, the higher the correlation degree is. The calculation method is shown in formula (10):
[0031]
[0032] In the formula, C(l) is the autocorrelation coefficient, l is the lag time point, q t represents the load current value at time t, q t+p represents the load current value at time t+p, is the mean of the entire load current sequence, and m is the total time.
[0033] Preferably, the load current change in step 2 may be related to multiple influencing factors. In order to fully consider the influence of other variable factors, the maximum information coefficient (MIC) calculation method is used to measure the linear and nonlinear correlation strength between two different variables. By giving two random variables with discrete values, x={x i ,i=1,2,…,n} and y={y i ,i=1,2,…,n}, where n is the number of elements in these variables, the mutual information I(x,y) is defined as:
[0034]
[0035] The mutual information is standardized and the maximum information coefficient is obtained as follows:
[0036]
[0037] Where p(x,y) is the joint probability between variables x and y, a and b are the number of grids divided in the x and y axis directions, B(n) is the upper limit of the number of grids divided, and the MIC value range is [0,1]. The larger the value, the higher the correlation between the load current and the external variable factors.
[0038] Preferably, in step 3, the CNN-GRU hybrid model decouples the two tasks of current quality high-level feature extraction and current quality time series prediction to achieve efficient processing and accurate prediction under current disturbance conditions. The current prediction model construction includes the determination of parameters such as the number of CNN layers, the number and size of convolution kernels, the number of GRU layers, and the number of GRU hidden units. First, the CNN-GRU model parameters are initialized, and the parameters of the GRU module are not changed during training. The MAE and RMSE after the number of CNN layers, the number and size of convolution kernels are analyzed by enumeration method to determine the CNN module parameters with the lowest prediction error; then, the CNN module parameters are not changed to conduct experiments, and the MAE and RMSE after the number of GRU layers and the number of hidden units are changed are calculated, and finally the optimal CNN-GRU model structure parameters are obtained.
[0039] Preferably, in step 2, the load current sequence respectively selects appropriate error evaluation indicators and variance evaluation indicators to measure the performance of the model in terms of prediction accuracy and robustness. The mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the performance of the model according to the magnitude and variation characteristics of the current data, which can better reflect the advantages and disadvantages of the model. The calculation method is shown in formula (13) and formula (14):
[0040]
[0041] Where n is the total number of predicted current samples, To predict the actual current value at the i-th time point, is the predicted current value at time point i.
[0042] Preferably, the arc current composite control in step 4 is to integrate the predicted value into the traditional PI control to form a composite control, and the input-output relationship of the controller is shown in formula (15):
[0043]
[0044] Where u(t) is the controller output value at time t, K is the controller gain, 1 / s is the integral function, and T i is the integration time constant, I ref is the current given value, I act is the actual value of current, I pre is the current prediction value, α is the prediction value weighting coefficient, and the value is in the interval [0,1].
[0045] (III) Beneficial effects
[0046] Compared with the prior art, the present invention provides a fast and accurate arc current control method based on plasma load current situation awareness, which has the following beneficial effects:
[0047] This fast and accurate arc current control method based on plasma load current situation perception uses a load current situation perception algorithm to accurately predict the load current change trend, thereby accurately regulating the load current in advance to achieve the purpose of eliminating the arc as early as possible, avoiding the power supply from entering the protection state, ensuring the constant load current during the coating process, and improving the quality of the deposited film layer. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 This is a working principle diagram of the fast and accurate arc current control based on plasma load current situation awareness proposed by the present invention;
[0049] Figure 2 This is a flow chart of the load current situation prediction and arc current control proposed by the present invention. DETAILED DESCRIPTION
[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0051] A working principle diagram of fast and accurate arc current control based on plasma load current situation awareness, such as Figure 1 As shown in the figure, the method mainly consists of two parts: plasma load current situation awareness and arc current composite control. The first part: constructing a plasma load current situation awareness model, inputting the sampled key parameter data, and outputting the load current prediction value; the second part: arc current composite control. The predicted value obtained by the load current situation awareness method is integrated into the traditional PI control to form a composite control, and then the control signal is output to the coating power supply, so as to achieve fast and accurate control of the arc current.
[0052] like Figure 2As described above, the load current situation prediction and arc current control process are mainly divided into four parts. The first part is the load current situation extraction. By collecting three electrical signals such as load current, voltage and frequency, as well as key parameters of the coating equipment such as chamber pressure, gas flow and ambient temperature, an information time series is constructed; the second part is the load current situation understanding. Through the autocorrelation analysis of safety features and multivariate correlation analysis, the load current change law is fully understood; the third part is the load current situation prediction. The CNN-GRU model algorithm is used to learn the high-level features of the data after situation understanding, and the long-term relationship between the series is mined to realize the prediction of the load current change trend; the fourth part is arc current control. The predicted value is integrated with the traditional proportional integral control to form a composite control with a feedforward function, and the current spike is suppressed through the load current feedforward to improve the quality of the deposited film layer.
[0053] This embodiment provides a method for fast and accurate arc current control based on plasma load current situation awareness, comprising the following steps:
[0054] Step 1: Load current factor extraction
[0055] In the vacuum coating deposition operation environment, the load operation status information is collected, including real-time current data, voltage data, frequency data, cavity pressure, gas flow and ambient temperature information. The time series is constructed based on this information, which is the data basis for realizing current situation awareness.
[0056] The switch power supply of the vacuum coating is equipped with a sensor unit and a data communication module with the coating equipment to construct the time series of three electrical signals: voltage U, current I, and frequency F at time t, and the time series of key parameters of the coating equipment such as the pressure P in the chamber, the gas flow Q, and the ambient temperature K:
[0057] V t =[v1,v2,…,v j ,…] Τ ,j∈Ν (1)
[0058] I t =[i1,i2,…,i j ,…] Τ ,j∈Ν (2)
[0059] F t =[f1,f2,…,f j ,…] Τ ,j∈Ν (3)
[0060] P t =[p1,p2,…,p j ,…] Τ ,j∈Ν (4)
[0061] Q t =[q1,q2,…,q j ,…] Τ ,j∈Ν (5)
[0062] K t =[k1,k2,…,k j ,…] Τ ,j∈Ν (6)
[0063] Since the sampling time of the sensor electrical signal is not synchronized with the data time of the key parameters of the coating equipment, the time scales of equations (1) to (6) are inconsistent. In order to avoid the above situation, the average value of the collected data is extracted at a certain interval time as a reference value according to the situation prediction time scale. Taking the time period current I and pressure P sequence as an example, the calculation method is shown in equations (7) to (8):
[0064]
[0065] In formula (7) and formula (8), i j1 and p j1 Corresponding to time period t j-1 +Δt current and pressure, i jm and p jm Corresponding to time period t j +Δt~t j +mΔt, and on this basis, the average current and pressure at time j are obtained, thereby deriving the current and pressure input time series of the model.
[0066] Step 2: Understanding the load current situation
[0067] Deeply mining the implicit correlation and data particularity in the data information provided by the load current element extraction, and fully understanding the load current variation law, is a deep understanding of the extracted information, which requires processing of abnormal values. First, an unsupervised learning method is used to detect abnormal values, including data density and cluster analysis, that is, according to the characteristics of the time series data itself, the load current sequence data density and the distance between the load current sequence data are used to distinguish normal values and abnormal values; secondly, for the processing of abnormal collected values, the continuity of the load current data and external factor data in the time series is ensured, and the missing values are filled by the local weighted regression interpolation method, while the abnormal deviation values are first removed and then locally weighted regression interpolation is performed;
[0068] To normalize the values of the load current data and external factor data, the calculation is as follows:
[0069]
[0070] In the formula, qt' represents the standardized data at time t, q t represents the original data at time t, q max and q min Respectively represent the maximum and minimum values in the time series data of each dimension;
[0071] As a kind of time series data, the load current data needs to be analyzed firstly in terms of its time series correlation, so as to provide the input time step of the situation prediction model to obtain the best prediction effect, that is, the historical data within a certain number of seconds from the current moment is used as input. The autocorrelation coefficient C(l) can measure the correlation degree of the load current time series data at two different time points. The value range of C(l) is [-1, 1]. If C(l)>0, the two load current time series data are positively correlated, otherwise they are negatively correlated. The larger |C(l)| is, the higher the correlation degree is. The calculation method is shown in formula (10):
[0072]
[0073] In the formula, C(l) is the autocorrelation coefficient, l is the lag time point, q t represents the load current value at time t, q t+p represents the load current value at time t+p, is the mean of the entire load current sequence, and m is the total time;
[0074] The load current change may be related to a variety of influencing factors. In order to fully consider the influence of other variable factors, the maximum information coefficient (MIC) calculation method is used to measure the linear and nonlinear correlation strength between two different variables. By giving two random variables with discrete values, x = {x i ,i=1,2,…,n} and y={y i ,i=1,2,…,n}, where n is the number of elements in these variables, the mutual information I(x,y) is defined as:
[0075]
[0076] The mutual information is standardized and the maximum information coefficient is obtained as follows:
[0077]
[0078] Where p(x,y) is the joint probability between variables x and y, a and b are the number of grids divided in the x and y axis directions, B(n) is the upper limit of the number of grids divided, and the MIC value interval is [0,1]. The larger the value, the higher the correlation between the load current and the external variable factors. The load current sequence selects appropriate error evaluation indicators and variance evaluation indicators to measure the performance of the model in terms of prediction accuracy and robustness. According to the order of magnitude and change characteristics of the current data, the mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the performance of the model, which can better reflect the advantages and disadvantages of the model. The calculation method is shown in formula (13) and formula (14):
[0079]
[0080] Where n is the total number of predicted current samples, To predict the actual current value at the i-th time point, is the predicted current value at time point i.
[0081] The arc current composite control in step 4 is to integrate the predicted value into the traditional PI control to form a composite control. The input-output relationship of the controller is shown in formula (15):
[0082]
[0083] Where u(t) is the controller output value at time t, K is the controller gain, 1 / s is the integral function, and T i is the integration time constant, I ref is the current given value, I act is the actual value of current, I pre is the current prediction value, α is the prediction value weighting coefficient, and the value is in the interval [0,1];
[0084] Step 3: Load current situation prediction
[0085] The CNN-GRU model algorithm is used to learn the high-level features of the data after situation understanding, to mine the relationship between long time series, and to predict the load current change trend. The CNN-GRU hybrid model decouples the two tasks of current quality high-level feature extraction and current quality time series prediction, and realizes efficient processing and accurate prediction under current disturbance. The current prediction model construction includes the determination of parameters such as the number of CNN layers, the number and size of convolution kernels, the number of GRU layers, and the number of GRU hidden units. First, the CNN-GRU model parameters are initialized, and the parameters of the GRU module are not changed during training. The MAE and RMSE of the CNN layer number, the number and size of convolution kernels are analyzed by enumeration method to determine the CNN module parameters with the lowest prediction error. Then, the CNN module parameters are not changed to conduct experiments, and the MAE and RMSE of the GRU layer number and the number of hidden units are calculated. Finally, the optimal CNN-GRU model structure parameters are obtained.
[0086] Step 4: Fast and accurate arc current control strategy
[0087] The predicted value is integrated with the traditional proportional-integral control to form a composite control with a feedforward function. The current spike is suppressed through load current feedforward to prevent the power supply from entering the protection state, thereby ensuring the constant load current in the process and improving the quality of the deposited film.
[0088] In summary, the fast and accurate arc current control method based on plasma load current situation awareness has the following specific implementation steps:
[0089] Step 1: Load current element extraction
[0090] The three electrical signals of voltage U, current I and frequency F are collected through the coating power supply sensor, and multiple data such as the key parameters of the coating equipment such as the pressure P in the coating equipment cavity, gas flow Q and ambient temperature K are collected through the communication module.
[0091] Step 2: Understanding the load current situation
[0092] 1) Processing of outliers. First, an unsupervised learning method is used to detect outliers, including data density and cluster analysis. That is, the normal values and outliers are distinguished according to the characteristics of the time series data itself, using the load current sequence data density and the distance between the load current sequence data; secondly, for the processing of abnormal collected values, in order to ensure the continuity of the load current data and external factor data in the time series, the local weighted regression interpolation method is a more effective method to fill in the missing values. For abnormal deviation values, they are first removed and then locally weighted regression interpolated.
[0093] 2) Standardized value processing: load current data and external factor data are standardized and calculated as follows:
[0094]
[0095] In the formula, q t ' represents the standardized data of the value at time t, q t represents the original data at time t, q max and q min Respectively represent the maximum and minimum values in the time series data of each dimension.
[0096] 3) Autocorrelation analysis requires the correlation of the load current analysis in time series to provide the input time step of the situation prediction model to obtain the best prediction effect, that is, the historical data within a certain number of seconds from the current moment is used as input. The autocorrelation coefficient C(l) can measure the correlation degree of the load current time series data at two different time points. The value range of C(l) is [-1, 1]. If C(l)>0, the two load current time series data are positively correlated, otherwise they are negatively correlated; the larger |C(l)| is, the higher the correlation degree is. The calculation method is as follows:
[0097]
[0098] Where: C(l) is the autocorrelation coefficient, l is the lag time point, q t represents the load current value at time t, q t+p represents the load current value at time t+p, is the mean of the entire load current sequence, and m is the total time.
[0099] 4) Maximum Information Coefficient Calculation, the load current change may be related to a variety of influencing factors. In order to fully consider the influence of other variable factors, the Maximum Information Coefficient (MIC) calculation method is used to measure the linear and nonlinear correlation strength between two different variables. By giving two random variables with discrete values, x = {x i ,i=1,2,…,n} and y={y i ,i=1,2,…,n}, where n is the number of elements in these variables, the mutual information I(x,y) is defined as:
[0100]
[0101] The mutual information is standardized and the maximum information coefficient is obtained as follows:
[0102]
[0103] Where p(x,y) is the joint probability between variables x and y, a and b are the number of grids divided in the x and y axis directions, B(n) is the upper limit of the number of grids divided, and the MIC value range is [0,1]. The larger the value, the higher the correlation between the load current and the external variable factors.
[0104] Step 3: Load current situation prediction
[0105] 1) Current prediction model based on CNN-GRU hybrid modeling. The hybrid model decouples the two tasks of current quality high-level feature extraction and current quality time series prediction to achieve efficient processing and accurate prediction under current disturbance. The construction of the current prediction model includes the determination of parameters such as the number of CNN layers, the number and size of convolution kernels, the number of GRU layers, and the number of GRU hidden units. First, the parameters of the CNN-GRU model are initialized. The parameters of the GRU module are not changed during training. The MAE and RMSE after the number of CNN layers, the number of convolution kernels, and the size are analyzed by enumeration method to determine the CNN module parameters with the lowest prediction error. Then, the experiments are conducted without changing the CNN module parameters, and the MAE and RMSE after the number of GRU layers and the number of hidden units are calculated. Finally, the optimal CNN-GRU model structure parameters are obtained.
[0106] 2) Evaluation of prediction results: Select appropriate error evaluation indicators and variance evaluation indicators for the load current sequence to measure the performance of the model in terms of prediction accuracy and robustness. Based on the magnitude and variation characteristics of the current data, the mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the performance of the model, which can better reflect the pros and cons of the model. The calculation method is:
[0107]
[0108] Where: n is the total number of predicted current samples, To predict the actual current value at the i-th time point, is the predicted current value at time point i.
[0109] Step 4: Fast and accurate arc current control strategy
[0110] The predicted value is integrated with the traditional proportional integral control to form a composite control with feedforward function. The input-output relationship of the controller is:
[0111]
[0112] Where u(t) is the controller output value at time t, K is the controller gain, 1 / s is the integral function, and T i is the integration time constant, Iref is the current given value, I act is the actual value of current, I pre is the current prediction value, α is the prediction value weighting coefficient, and the value is in the interval [0,1].
[0113] It should be noted that the term "comprises" or any other variation thereof is intended to cover non-exclusive inclusion, so that a process, method, article, or device that includes a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such process, method, article, or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article, or device that includes the element.
[0114] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A fast and accurate arc current control method based on plasma load current situation awareness, It is characterized in that The following steps are involved: Step 1: Load current factor extraction In the vacuum coating deposition operation environment, the load operation status information is collected, including real-time current data, voltage data, frequency data, cavity pressure, gas flow and ambient temperature information. The time series is constructed based on this information, which is the data basis for realizing current situation awareness. Step 2: Understanding the load current situation Deeply explore the implicit correlation and data specificity in the data information provided by the load current element extraction, and fully understand the load current variation law, which is a deep understanding of the extracted information; Step 3: Load current situation prediction The CNN-GRU model algorithm is used to learn the high-level features of the data after situation understanding, to mine the before and after relationships of long time series, and to predict the trend of load current changes; Step 4: Fast and accurate arc current control strategy The predicted value is integrated with the traditional proportional-integral control to form a composite control with a feedforward function. The current spike is suppressed through load current feedforward to prevent the power supply from entering the protection state, thereby ensuring the constant load current in the process and improving the quality of the deposited film.
2. A fast and accurate arc current control method based on plasma load current situation awareness according to claim 1, Features: The switch power supply of the vacuum coating in step 1 is equipped with a sensor unit, and a data communication module with the coating equipment, to construct the time series of three electrical signals of voltage U, current I, and frequency F at time t, and the time series of key parameters of the coating equipment such as the pressure P in the cavity, the gas flow rate Q, and the ambient temperature K: V t =[v 1 ,v 2 ,…,v j ,…] Τ ,j∈Ν (1) I t =[i 1 ,i 2 ,…,i j ,…] Τ ,j∈Ν (2) F t =[f 1 ,f 2 ,…,f j ,…] Τ ,j∈Ν (3) P t =[p 1 ,p 2 ,…,p j ,…] Τ ,j∈Ν (4) Q t =[q 1 ,q 2 ,…,q j ,…] Τ ,j∈Ν (5) K t =[k 1 ,k 2 ,…,k j ,…] Τ ,j∈Ν (6) Since the sampling time of the sensor electrical signal is not synchronized with the data time of the key parameters of the coating equipment, the time scales of equations (1) to (6) are inconsistent. In order to avoid the above situation, the average value of the collected data is extracted at a certain interval time as a reference value according to the situation prediction time scale. Taking the time period current I and pressure P sequence as an example, the calculation method is shown in equations (7) to (8): In formula (7) and formula (8), i j1 and p j1 Corresponding to time period t j-1 +Δt current and pressure, i jm and p jm Corresponding to time period t j +Δt~t j +mΔt, and on this basis, the average current and pressure at time j are obtained, thereby deriving the current and pressure input time series of the model.
3. According to claim 1, a fast and accurate arc current control method based on plasma load current situation awareness, Features: In step 2, outliers need to be processed. First, an unsupervised learning method is used to detect outliers, including data density and cluster analysis, that is, normal values and outliers are distinguished based on the characteristics of the time series data itself, using the load current sequence data density and the distance between the load current sequence data; secondly, for the processing of abnormal collected values, the continuity of the load current data and the external factor data in the time series is ensured, and the missing values are filled by the local weighted regression interpolation method, while the abnormal deviation values are first removed and then locally weighted regression interpolation is performed.
4. According to claim 1, a fast and accurate arc current control method based on plasma load current situation awareness, Features: In step 2, the load current data and external factor data are to be normalized, and the calculation is as follows: In the formula, q t ' represents the standardized data of the value at time t, q t represents the original data at time t, q max and q min Respectively represent the maximum and minimum values in the time series data of each dimension.
5. A fast and accurate arc current control method based on plasma load current situation awareness according to claim 1, Features: In step 2, the load current data is a kind of time series data. First, its time series correlation needs to be analyzed to provide the input time step of the situation prediction model to obtain the best prediction effect, that is, the historical data within a certain number of seconds from the current moment is used as input. The autocorrelation coefficient C(l) can measure the correlation degree of the load current time series data at two different time points. The value range of C(l) is [-1, 1]. If C(l)>0, the two load current time series data are positively correlated, otherwise they are negatively correlated. The larger |C(l)| is, the higher the correlation degree is. The calculation method is shown in formula (10): In the formula, C(l) is the autocorrelation coefficient, l is the lag time point, q t represents the load current value at time t, q t+p represents the load current value at time t+p, is the mean of the entire load current sequence, and m is the total time.
6. A fast and accurate arc current control method based on plasma load current situation awareness according to claim 1, Features: The load current change in step 2 may be related to a variety of influencing factors. In order to fully consider the influence of other variable factors, the maximum information coefficient (MIC) calculation method is used to measure the linear and nonlinear correlation strength between two different variables. By giving two random variables with discrete values, x={x i ,i=1,2,…,n} and y={y i ,i=1,2,…,n}, where n is the number of elements in these variables, the mutual information I(x,y) is defined as: The mutual information is standardized and the maximum information coefficient is obtained as follows: Where p(x,y) is the joint probability between variables x and y, a and b are the number of grids divided in the x and y axis directions, B(n) is the upper limit of the number of grids divided, and the MIC value range is [0,1]. The larger the value, the higher the correlation between the load current and the external variable factors.
7. A fast and accurate arc current control method based on plasma load current situation awareness according to claim 1, Features: In step 3, the CNN-GRU hybrid model decouples the two tasks of current quality high-level feature extraction and current quality time series prediction to achieve efficient processing and accurate prediction under current disturbance conditions. The current prediction model construction includes the determination of parameters such as the number of CNN layers, the number and size of convolution kernels, the number of GRU layers, and the number of GRU hidden units. First, the CNN-GRU model parameters are initialized, and the parameters of the GRU module are not changed during training. The MAE and RMSE after the number of CNN layers, the number and size of convolution kernels are analyzed by enumeration method to determine the CNN module parameters with the lowest prediction error; then, the CNN module parameters are not changed to conduct experiments, and the MAE and RMSE after the number of GRU layers and the number of hidden units are changed are calculated, and finally the optimal CNN-GRU model structure parameters are obtained.
8. A fast and accurate arc current control method based on plasma load current situation awareness according to claim 1, Features: In step 2, the load current sequence selects appropriate error evaluation indicators and variance evaluation indicators to measure the performance of the model in terms of prediction accuracy and robustness. The mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the performance of the model according to the magnitude and variation characteristics of the current data, which can better reflect the advantages and disadvantages of the model. The calculation method is shown in formula (13) and formula (14): Where n is the total number of predicted current samples, To predict the actual current value at the i-th time point, is the predicted current value at time point i.
9. A fast and accurate arc current control method based on plasma load current situation awareness according to claim 1, Features: The arc current composite control in step 4 is to integrate the predicted value into the traditional PI control to form a composite control. The input-output relationship of the controller is shown in formula (15): Where u(t) is the controller output value at time t, K is the controller gain, 1 / s is the integral function, and T i is the integration time constant, I ref is the current given value, I act is the actual value of current, I pre is the current prediction value, α is the prediction value weighting coefficient, and the value is in the interval [0,1].
