Adaptive arima modeling method and system for semiconductor manufacturing equipment predictive maintenance

Abstract

The application relates to the technical field of predictive maintenance of semiconductor manufacturing equipment, and particularly discloses a self-adaptive ARIMA modeling method and system for predictive maintenance of semiconductor manufacturing equipment; the method comprises the following steps: collecting process parameter data in real time, such as etching rate, temperature or pressure; dynamically adjusting the ARIMA model training window size based on data volatility, automatically expanding or reducing the window to optimize model fitting by calculating the rolling standard deviation and comparing the rolling standard deviation with an adaptive threshold based on historical stable standard deviation; fitting the ARIMA model by using the adjusted window data; and predicting by using the model and generating a device state early warning signal. The application solves the prediction lag and false alarm problems caused by the fixed parameter ARIMA model failing to adapt to the dynamic degradation of the device, realizes early fault warning, and improves the device operation efficiency and product yield.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing technology, and in particular relates to an adaptive ARIMA modeling method and system for predictive maintenance of semiconductor manufacturing equipment. Background Technology

[0002] In semiconductor manufacturing, the process chambers of critical equipment such as etching, deposition, and photolithography experience performance degradation over time. This degradation can be caused by factors such as RF power attenuation, gas nozzle blockage, or deposit accumulation on chamber walls, leading to drift or abrupt changes in process parameters (e.g., etching rate, temperature, or pressure). Current technologies typically rely on periodic maintenance or statistical process control (SPC) threshold alarms. However, these methods suffer from latency, triggering alarms only when parameters significantly exceed control limits (e.g., ±3σ), by which time a large number of defective wafers may have already been produced. Furthermore, they have high false alarm rates and cannot effectively distinguish between normal slow drift and sudden anomalies, resulting in unnecessary equipment downtime. To further improve prediction accuracy, existing technologies also employ time-series models with fixed parameters (such as ARIMA models) for predictive control. However, the training window size and parameters of ARIMA models are often preset to fixed values, failing to adapt to the dynamic changes in process parameters during equipment degradation. When the equipment status changes, the prediction accuracy of the fixed-window ARIMA model gradually decreases due to outdated data or noise interference, making it difficult to achieve early fault warning. This has become a major technical problem in predictive maintenance of semiconductor manufacturing equipment. Summary of the Invention

[0003] This invention aims to address the problem of inaccurate predictions caused by the inability of fixed-parameter models to adapt to dynamic changes by adaptively adjusting the training window of the ARIMA model. To overcome the shortcomings of the prior art, this invention provides an adaptive ARIMA modeling method, system, device, and medium for predictive maintenance of semiconductor manufacturing equipment. By introducing a dynamic window adjustment mechanism based on the volatility of process parameter data, it significantly improves the adaptability and accuracy of the ARIMA model in predictive maintenance of semiconductor manufacturing equipment.

[0004] The first aspect of this invention discloses an adaptive ARIMA modeling method for predictive maintenance of semiconductor manufacturing equipment, comprising the following steps:

[0005] Step S101: Real-time acquisition of process parameter data of semiconductor manufacturing equipment, wherein the process parameter data includes at least one of etching rate, temperature or pressure;

[0006] Step S102: Based on the volatility of the process parameter data, dynamically adjust the training window size of the ARIMA model, wherein the dynamic adjustment includes:

[0007] Calculate the rolling standard deviation σ of recent dataroll The rolling standard deviation σ roll It is the standard deviation of the most recent m data points, where m is a preset positive integer and the value of m ranges from 5 to 20;

[0008] σ roll With upper limit threshold Th high and lower limit threshold Th low Comparison, where Th high =k1×σ process ,Th low =k2×σ process k1 and k2 are preset coefficients greater than 0, with k1 ranging from 0.8 to 1.5 and k2 ranging from 0.1 to 0.5. σ process The historical stability standard deviation of the process parameters during the stable operation phase of the equipment;

[0009] Based on the rolling standard deviation σ roll With upper limit threshold Th high and lower limit threshold Th low Adjusting the training window size based on the comparison results includes:

[0010] If σ roll >Th high Then increase the training window size to min(W) old ×a,W max ), where W old The current window size is denoted by 'a', where 'a' is a preset coefficient greater than 1, with a value ranging from 1.05 to 1.2. W max To preset the maximum window size, W max The value range is from 200 to 500;

[0011] If σ roll <Th low Then reduce the training window size to max(W) old ×b,W min ), where b is a preset coefficient less than 1, and the value of b ranges from 0.8 to 0.95, W min To preset the minimum window size, W min The value range is from 20 to 50;

[0012] If Th low ≤σ roll ≤Th high If so, the training window size remains unchanged;

[0013] Step S103: Using the process parameter data within the adjusted training window, fit the ARIMA model, wherein the ARIMA model includes the autoregressive order p, the difference order d, and the moving average order q;

[0014] Step S104: Use the fitted ARIMA model to predict the next process parameter data, and generate equipment status warning signals based on the prediction results.

[0015] The above adaptive ARIMA modeling method, after step S103, further includes step S105: monitoring changes in ARIMA model parameters, wherein monitoring includes:

[0016] Calculate the difference order d of the currently fitted ARIMA model. current The difference order d between the previous fitted model and the previous fitted model previous The difference Δd = |d current -d previous |;

[0017] Compare Δd with the control upper limit UCL, where UCL = c × σ delta c is a preset coefficient greater than 0, and the value of c ranges from 2 to 4. σ delta The standard deviation of the historical d-value change sequence during the equipment health and stability phase is defined by the time period in historical data when the fluctuation of process parameters is less than a preset threshold.

[0018] If Δd > UCL, it is determined to be a parameter mutation, and a parameter mutation signal is generated.

[0019] The adaptive ARIMA modeling method described above, in step S105, further includes: if in three consecutive calculations, Δd changes in the same direction, and each Δd is greater than k×σ. delta , where k is a preset coefficient, and the value of k ranges from 1 to 2, then it is determined to be a parameter mutation, and a parameter mutation signal is generated.

[0020] In the aforementioned adaptive ARIMA modeling method, the parameter mutation signal is used to trigger fault diagnosis, wherein the fault diagnosis includes:

[0021] Analyze the equipment operation data corresponding to the parameter mutation signal and check whether there are abnormal fluctuations in other process parameters during the same period;

[0022] Based on the analysis results, the type of equipment degradation is correlated to determine the root cause.

[0023] The aforementioned adaptive ARIMA modeling method, after step S104, further includes step S106: analyzing the statistical characteristics of the predicted residuals and matching them with a failure mode library, wherein the analysis includes:

[0024] Calculate the residual between the predicted value and the actual value, and calculate the statistical characteristics of the residual, including at least one of variance, mean, skewness or kurtosis;

[0025] The statistical features are compared with the mapping rules in the preset fault mode library, which is trained based on historical fault data, and the mapping rules include the correspondence between fault types and residual statistical features.

[0026] Based on the matching results, determine the fault type and alarm level.

[0027] In the above adaptive ARIMA modeling method, step S103, fitting the ARIMA model includes:

[0028] The stationarity of the process parameter data within the training window is tested using the extended Dickey-Fuller test. If the p-value is greater than 0.05, the difference operation is performed until the sequence is stationary. The difference order is denoted as d.

[0029] For stationary sequences, the optimal autoregression order p and moving average order q can be determined through autocorrelation function and partial autocorrelation function analysis, or by using the auto_arima algorithm for grid search.

[0030] In the aforementioned adaptive ARIMA modeling method, the equipment status warning signal includes a warning level and a severity level. After generating the equipment status warning signal, it is sent to the equipment maintenance system or engineer terminal through an interface.

[0031] A second aspect of this invention discloses an adaptive ARIMA modeling system for predictive maintenance of semiconductor manufacturing equipment, comprising:

[0032] The data acquisition module is configured to acquire process parameter data of semiconductor manufacturing equipment in real time, wherein the process parameter data includes at least one of etching rate, temperature or pressure.

[0033] A window adjustment module is configured to dynamically adjust the training window size of the ARIMA model based on the volatility of the process parameter data. The window adjustment module includes:

[0034] The volatility calculation unit is configured to calculate the rolling standard deviation σ of recent data. roll The rolling standard deviation σ roll It is the standard deviation of the most recent m data points, where m is a preset positive integer;

[0035] The threshold comparison unit is configured to compare σ roll With upper limit threshold Th high and lower limit threshold Th low Comparison, where Th high =k1×σ process ,Th low =k2×σ process k1 and k2 are preset coefficients greater than 0, σ processThe historical stability standard deviation of the process parameters during the stable operation phase of the equipment;

[0036] Window control unit, configured to operate based on rolling standard deviation σ roll With upper limit threshold Th high and lower limit threshold Th low Adjusting the training window size based on the comparison results includes:

[0037] If σ roll >Th high Then increase the training window size to min(W) old ×a,W max ), where W old Where W is the current window size, 'a' is a preset coefficient greater than 1, and W is the current window size. max Set the maximum window size to the preset value;

[0038] If σ roll <Th low Then reduce the training window size to max(W) old ×b,W min ), where b is a preset coefficient less than 1, W min Set the minimum window size;

[0039] If Th low ≤σ roll ≤Th high If so, the training window size remains unchanged;

[0040] The model fitting module is configured to use the process parameter data within the adjusted training window to fit an ARIMA model, wherein the ARIMA model includes an autoregressive order p, a difference order d, and a moving average order q.

[0041] The prediction and early warning module is configured to use the fitted ARIMA model to predict the next process parameter data and generate equipment status early warning signals based on the prediction results.

[0042] A third aspect of the present invention discloses an electronic device, comprising: a memory and a processor, wherein the processor and the memory are connected;

[0043] The memory is used to store programs;

[0044] The processor invokes a program stored in the memory to execute the method provided by the first aspect embodiment and / or any possible embodiment in combination with the first aspect embodiment.

[0045] The fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a computer to perform the methods provided by the embodiments of the first aspect and / or any possible embodiments in combination with the embodiments of the first aspect.

[0046] Compared with existing technologies, this invention has the following advantages: by calculating the rolling standard deviation of recent data and comparing it with an adaptive threshold based on historical stable standard deviations, the training window size can be optimized in real time: when process fluctuations are large, the window is automatically expanded to smooth noise and enhance model stability; when fluctuations are small, the window is reduced to quickly respond to minor changes and improve model sensitivity. This adaptive adjustment effectively overcomes the problem of decreased prediction accuracy caused by the inability of fixed-parameter ARIMA models to adapt to the dynamic temporal characteristics during equipment degradation, thereby reducing alarm lag and false alarm rate in traditional methods, achieving early warning of equipment failures, and reducing unplanned downtime and yield loss.

[0047] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method in Example 1.

[0049] Figure 2 This is a system module diagram of Example 2. Detailed Implementation

[0050] Example 1

[0051] like Figure 1 As shown, an adaptive ARIMA modeling method for predictive maintenance of semiconductor manufacturing equipment includes the following steps:

[0052] Step S101: Real-time acquisition of process parameter data of semiconductor manufacturing equipment, wherein the process parameter data includes at least one of etching rate, temperature or pressure;

[0053] During implementation, sensors integrated within the plasma etching machine chamber continuously collect etching rate data, measured in angstroms per minute, using the completion of each wafer as a trigger point. The collected data is transmitted to a central database in real time, ensuring the integrity of the time series. For example, when the etching machine processes the 100th wafer, if the sensor measures an etching rate of 500 angstroms per minute, the system immediately records and stores this value. By acquiring process parameter data in real time, continuous monitoring of the equipment's operating status is achieved, providing an accurate data foundation for subsequent analysis.

[0054] Step S102: Based on the volatility of the process parameter data, dynamically adjust the training window size of the ARIMA model, wherein the dynamic adjustment includes:

[0055] Calculate the rolling standard deviation σ of recent data roll The rolling standard deviation σ roll It is the standard deviation of the most recent m data points, where m is a preset positive integer and the value of m ranges from 5 to 20;

[0056] σ roll With upper limit threshold Th high and lower limit threshold Th low Comparison, where Th high =k1×σ process ,Th low =k2×σ process k1 and k2 are preset coefficients greater than 0, with k1 ranging from 0.8 to 1.5 and k2 ranging from 0.1 to 0.5. σ process The historical stability standard deviation of the process parameters during the stable operation phase of the equipment;

[0057] Based on the rolling standard deviation σ roll With upper limit threshold Th high and lower limit threshold Th low Adjusting the training window size based on the comparison results includes:

[0058] If σ roll >Th high Then increase the training window size to min(W) old ×a,W max ), where W old The current window size is denoted by 'a', where 'a' is a preset coefficient greater than 1, with a value ranging from 1.05 to 1.2. W max To preset the maximum window size, W max The value range is from 200 to 500;

[0059] If σ roll <Th low Then reduce the training window size to max(W) old ×b,W min ), where b is a preset coefficient less than 1, and the value of b ranges from 0.8 to 0.95, W min To preset the minimum window size, W min The value range is from 20 to 50;

[0060] If Th low ≤σ roll ≤Th high If so, the training window size remains unchanged;

[0061] The implementation includes three sub-steps: First, calculate the rolling standard deviation σ of recent data. rollThe rolling window size m is set to 10, meaning the standard deviation is calculated using the 10 most recent etch rate data points to reflect short-term volatility. For example, the system extracts the etch rate values ​​from wafers 91 to 100 from the database and calculates σ. roll It is 0.6 angstroms per minute. Secondly, σ... roll With upper limit threshold Th high and lower limit threshold Th low Compare; The high By multiplying the coefficient k1=1.0 by the historical stable standard deviation σ process Calculations show that Th low By multiplying σ by the coefficient k2=0.3 process The calculation yields σ. process Based on historical data from the stable operation phase of the equipment, the speed was determined to be 0.5 angstroms per minute. Therefore, Th high Th is 0.5 low The value is 0.15. Finally, the training window size is adjusted based on the comparison results: if σ roll Greater than Th high Then expand the window to the current window size W. old The value after multiplying by the coefficient a=1.1, and not exceeding the maximum window size W. max =300; if σ roll Less than Th low Then shrink the window to W old The value after multiplying by the coefficient b=0.9, and not less than the minimum window size Wmin=30; if σroll is within the threshold range, the window remains unchanged. For example, when σ... roll =0.6 greater than Th high When the value is 0.5, the system expands the training window from 100 wafers to 110 wafers. By dynamically adjusting the training window size, an adaptive match is achieved between the amount of model training data and the current process volatility. This allows for smoothing noise and improving stability when data fluctuations are large, while reducing the window size enhances sensitivity to minute changes when data is stable.

[0062] Step S103: Using the process parameter data within the adjusted training window, fit the ARIMA model, wherein the ARIMA model includes the autoregressive order p, the difference order d, and the moving average order q;

[0063] During implementation, the etching rate data sequence within the current training window is extracted from the database, for example, 110 data points. First, a stationarity test is performed: the extended Dickey-Fuller test is applied to calculate the p-value. If the p-value is greater than 0.05, the sequence is first-order differencing is performed, and the test is repeated until the p-value is less than 0.05. The differencing order is denoted as d. For example, the initial sequence has a p-value of 0.8, indicating non-stationarity; after first-order differencing, the p-value is 0.01, indicating stationarity, therefore d=1. Next, on the stationary sequence, the optimal autoregression order p and moving average order q are determined by analyzing the truncation characteristics of the autocorrelation function and partial autocorrelation function, or by using the Auto_Arim algorithm for grid search, with the goal of minimizing the Akaike information criterion. For example, after analysis, p=1 and q=1 are determined, thus fitting an ARIMA(1,1,1) model. By fitting the ARIMA model, accurate capture of the time-series dynamics of process parameters is achieved, providing a reliable basic model for prediction.

[0064] Step S104: Use the fitted ARIMA model to predict the next process parameter data, and generate equipment status warning signals based on the prediction results.

[0065] During implementation, a fitted model is used to predict the etching rate of the next wafer; for example, the predicted value for the 111th wafer is 505 angstroms per minute. After actual processing is completed, the actual value is recorded and the predicted residual is calculated. If the residual exceeds a threshold (e.g., ±2 standard deviations) set based on the historical residual standard deviation, an equipment status warning signal is generated, such as a warning level signal, and sent to the equipment maintenance system via the application programming interface. For example, if the actual etching rate is 510 angstroms per minute and the residual is 5 angstroms per minute, exceeding the threshold, the system triggers a warning signal. Through real-time prediction and residual monitoring, early warning of potential equipment failures is achieved, thereby reducing unplanned downtime and improving production efficiency.

[0066] In one embodiment, after step S103, a step S105 is further included: monitoring changes in ARIMA model parameters, wherein the monitoring includes:

[0067] Calculate the difference order d of the currently fitted ARIMA model. current The difference order d between the previous fitted model and the previous fitted model previous The difference Δd = |d current -d previous |;

[0068] Compare Δd with the control upper limit UCL, where UCL = c × σ delta c is a preset coefficient greater than 0, and the value of c ranges from 2 to 4. σ deltaThe standard deviation of the historical d-value change sequence during the equipment health and stability phase is defined by the time period in historical data when the fluctuation of process parameters is less than a preset threshold.

[0069] If Δd > UCL, it is determined to be a parameter mutation, and a parameter mutation signal is generated.

[0070] During implementation, after each fitting of a new ARIMA model, the current difference order d is extracted from the model parameters. current And the difference order d with the previous model previous Compare and calculate the absolute value difference Δd = |d current - d previous For example, the previous model d previous =1, current model d current =2, then Δd=1. The upper limit of control (UCL) is calculated by multiplying the coefficient c=3 by the standard deviation σ of the historical d values. delta Calculation yielded σ delta Based on historical data from the equipment's healthy and stable phases, for example, by systematically analyzing the change sequence of the d-value during the processing of the past 1000 wafers, σ can be calculated. delta =0.2, therefore UCL=0.6. If Δd>UCL, it is determined to be a parameter mutation, and a parameter mutation signal is generated. By monitoring the mutation of the differential order d, early identification of step degradation in equipment performance is achieved, thereby triggering early warning in situations such as accelerated deposition on the chamber wall and avoiding batch wafer loss.

[0071] In one embodiment, step S105 further includes: if in three consecutive calculations, Δd changes in the same direction, and each Δd is greater than k×σ delta , where k is a preset coefficient, and the value of k ranges from 1 to 2, then it is determined to be a parameter mutation, and a parameter mutation signal is generated.

[0072] During implementation, record the Δd values ​​for three consecutive model fits and check if they all change in the same direction. For example, if the three consecutive Δd values ​​are 0.3, 0.4, and 0.5, they all increase positively. Simultaneously, each Δd must be greater than k×σ. delta Where k is set to 1.5, σ delta =0.2, therefore the threshold is 0.3; since 0.3, 0.4, and 0.5 are all greater than 0.3 and change in the same direction, the system determines it as a parameter mutation and generates a signal. Through continuous trend analysis, reliable detection of gradual degradation is achieved, false alarms caused by single data fluctuations are reduced, and the stability of mutation determination is improved.

[0073] In one embodiment, the parameter mutation signal is used to trigger fault diagnosis, wherein the fault diagnosis includes:

[0074] Analyze the equipment operation data corresponding to the parameter mutation signal and check whether there are abnormal fluctuations in other process parameters during the same period;

[0075] Based on the analysis results, the type of equipment degradation is correlated to determine the root cause.

[0076] During implementation, upon generating a parameter mutation signal, the system automatically correlates it with equipment operating data from the same period, checking for abnormal fluctuations in other process parameters such as RF power and gas flow. The system extracts data from 10 wafers before and after the mutation point from the database and calculates the rolling standard deviation of these parameters. If the standard deviation of RF power increases from 0.1 to 0.5, it correlates this with a pre-defined fault type, such as RF power degradation, based on the d-value mutation. Through multi-parameter collaborative analysis, the system enables rapid localization of the root cause of the fault, guiding maintenance personnel to clean or replace components in a targeted manner, thereby improving maintenance efficiency.

[0077] In one embodiment, after step S104, a step S106 is further included: analyzing the statistical characteristics of the predicted residuals and matching them with a failure mode library, wherein the analysis includes:

[0078] Calculate the residual between the predicted value and the actual value, and calculate the statistical characteristics of the residual, including at least one of variance, mean, skewness or kurtosis;

[0079] The statistical features are compared with the mapping rules in the preset fault mode library, which is trained based on historical fault data, and the mapping rules include the correspondence between fault types and residual statistical features.

[0080] Based on the matching results, determine the fault type and alarm level.

[0081] During implementation, the residual between the predicted and actual values ​​is calculated. For example, for the 111th wafer, the predicted etching rate is 505 Å per minute, the actual rate is 510 Å per minute, and the residual is 5 Å per minute. Then, the statistical characteristics of the residual sequence are calculated, including variance, mean, skewness, and kurtosis. For example, the variance of the last 20 residuals increases from 0.1 to 0.5, and the mean shows a positive shift of 0.3. The fault mode library is trained based on historical fault data and includes mapping rules such as increased variance corresponding to "gas nozzle blockage" and a positive shift in the mean corresponding to "chamber wall deposition." The system compares the calculated features with the rules in the library. If a match is found, the fault type is determined to be gas nozzle blockage, and a warning level alarm is triggered. Through residual feature analysis, refined fault type identification is achieved, thereby providing specific maintenance recommendations and reducing unnecessary equipment downtime.

[0082] In one embodiment, fitting the ARIMA model in step S103 includes:

[0083] The stationarity of the process parameter data within the training window is tested using the extended Dickey-Fuller test. If the p-value is greater than 0.05, the difference operation is performed until the sequence is stationary. The difference order is denoted as d.

[0084] For stationary sequences, the optimal autoregression order p and moving average order q can be determined through autocorrelation function and partial autocorrelation function analysis, or by using the Auto_Arim algorithm for grid search.

[0085] In implementation, the stationarity of the etching rate data within the training window is tested using the extended Dickey-Fuller test to calculate the p-value. If the p-value is greater than 0.05, it indicates that the sequence is non-stationary, so first-order differencing is performed, and the test is repeated until the p-value is less than 0.05. The differencing order is denoted as d. For example, if the initial sequence p-value is 0.8, the p-value after first-order differencing is 0.01, therefore d=1. For stationary sequences, truncation features are analyzed using autocorrelation and partial autocorrelation functions, or a grid search is performed using the auto_arima algorithm, with the goal of minimizing the Akaike information criterion, to determine the optimal p and q values, such as p=1 and q=1. The standardized fitting process ensures the model's adaptability to non-stationary data, improving the accuracy and robustness of predictions.

[0086] In one embodiment, the device status warning signal includes a warning level and a severity level. After the device status warning signal is generated, it is sent to the device maintenance system or engineer terminal through an interface.

[0087] During implementation, alarm levels are categorized based on the degree of abrupt changes in predicted residuals or parameters; for example, residuals exceeding twice the historical residual standard deviation are set to a warning level, and those exceeding three times are set to a critical level. Signals are sent in real-time to the equipment maintenance system or engineer's terminal via a REST API, with the JSON format including timestamp, device ID, fault type, and level. Through tiered alarms and real-time transmission, priority management of maintenance responses is achieved, thereby optimizing resource allocation and shortening fault handling time.

[0088] Example 2

[0089] like Figure 2 As shown, an adaptive ARIMA modeling system for predictive maintenance of semiconductor manufacturing equipment includes:

[0090] The data acquisition module is configured to acquire process parameter data of semiconductor manufacturing equipment in real time, wherein the process parameter data includes at least one of etching rate, temperature or pressure.

[0091] A window adjustment module is configured to dynamically adjust the training window size of the ARIMA model based on the volatility of the process parameter data. The window adjustment module includes:

[0092] The volatility calculation unit is configured to calculate the rolling standard deviation σ of recent data. roll The rolling standard deviation σ roll It is the standard deviation of the most recent m data points, where m is a preset positive integer;

[0093] The threshold comparison unit is configured to compare σ roll With upper limit threshold Th high and lower limit threshold Th low Comparison, where Th high =k1×σ process ,Th low =k2×σ process k1 and k2 are preset coefficients greater than 0, σ process The historical stability standard deviation of the process parameters during the stable operation phase of the equipment;

[0094] Window control unit, configured to operate based on rolling standard deviation σ roll With upper limit threshold Th high and lower limit threshold Th low Adjusting the training window size based on the comparison results includes:

[0095] If σ roll >Th high Then increase the training window size to min(W) old ×a,W max ), where W old Where W is the current window size, 'a' is a preset coefficient greater than 1, and W is the current window size. max Set the maximum window size to the preset value;

[0096] If σ roll <Th low Then reduce the training window size to max(W) old ×b,W min ), where b is a preset coefficient less than 1, W min Set the minimum window size;

[0097] If Th low ≤σ roll ≤Th high If so, the training window size remains unchanged;

[0098] The model fitting module is configured to use the process parameter data within the adjusted training window to fit an ARIMA model, wherein the ARIMA model includes an autoregressive order p, a difference order d, and a moving average order q.

[0099] The prediction and early warning module is configured to use the fitted ARIMA model to predict the next process parameter data and generate equipment status early warning signals based on the prediction results.

[0100] The adaptive ARIMA modeling system for predictive maintenance of semiconductor manufacturing equipment provided in this embodiment has the same implementation principle and technical effect as the method embodiment in Embodiment 1. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in Embodiment 1.

[0101] Example 3

[0102] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a computer to perform the adaptive ARIMA modeling method for predictive maintenance of semiconductor manufacturing equipment as described in Embodiment 1 above.

[0103] Example 4

[0104] An electronic device includes: a memory and a processor, wherein the processor and the memory are connected;

[0105] The memory is used to store programs;

[0106] The processor invokes a program stored in the memory to execute an adaptive ARIMA modeling method for predictive maintenance of semiconductor manufacturing equipment as described in Embodiment 1.

[0107] It should be noted that the electronic device mentioned may be, but is not limited to, personal computers (PCs), tablet computers, mobile internet devices (MIDs), etc.

[0108] It should be noted that processors, memory, and other components that may be present in electronic devices are electrically connected to each other, directly or indirectly, to enable data transmission or interaction. For example, processors, memory, and other components may be electrically connected to each other via one or more communication buses or signal lines.

[0109] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0110] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0111] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0112] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, laptop, server, mobile phone, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An adaptive ARIMA modeling method for predictive maintenance of semiconductor manufacturing equipment, characterized in that, Includes the following steps: Step S101: Real-time acquisition of process parameter data of semiconductor manufacturing equipment, wherein the process parameter data includes at least one of etching rate, temperature or pressure; Step S102: Based on the volatility of the process parameter data, dynamically adjust the training window size of the ARIMA model, wherein the dynamic adjustment includes: Calculate the rolling standard deviation σ of recent data roll The rolling standard deviation σ roll It is the standard deviation of the most recent m data points, where m is a preset positive integer; σ roll With upper limit threshold Th high and lower limit threshold Th low Comparison, where Th high =k1×σ process ,Th low =k2×σ process k1 and k2 are preset coefficients greater than 0, σ process The historical stability standard deviation of the process parameters during the stable operation phase of the equipment; Based on the rolling standard deviation σ roll With upper limit threshold Th high and lower limit threshold Th low Adjusting the training window size based on the comparison results includes: If σ roll >Th high Then increase the training window size to min(W) old ×a,W max ), where W old Where W is the current window size, 'a' is a preset coefficient greater than 1, and W is the current window size. max Set the maximum window size to the preset value; If σ roll <Th low Then reduce the training window size to max(W) old ×b,W min ), where b is a preset coefficient less than 1, W min Set the minimum window size; If Th low ≤σ roll ≤Th high If so, the training window size remains unchanged; Step S103: Using the process parameter data within the adjusted training window, fit the ARIMA model, wherein the ARIMA model includes the autoregressive order p, the difference order d, and the moving average order q; Step S104: Use the fitted ARIMA model to predict the next process parameter data, and generate equipment status warning signals based on the prediction results; Generating an equipment status warning signal includes: using a fitted ARIMA model to predict the process parameters for the next process cycle and obtaining predicted values; after the actual completion of the process cycle, obtaining the actual values ​​of the process parameters and calculating the residual between the predicted and actual values; comparing this residual with a pre-set threshold, which is set based on the standard deviation of historical prediction residuals; if the residual exceeds the threshold, generating an equipment status warning signal. Following step S103, step S105 is also included: monitoring changes in ARIMA model parameters, wherein monitoring includes: Calculate the difference order d of the currently fitted ARIMA model. current The difference order d with the previous fitted model previous The difference Δd = |d current -d previous |; Compare Δd with the control upper limit UCL, where UCL = c × σ delta c is a preset coefficient greater than 0, σ delta The standard deviation of the historical d-value change sequence during the equipment health and stability phase is defined by the time period in historical data when the fluctuation of process parameters is less than a preset threshold. If Δd > UCL, it is determined to be a parameter mutation, and a parameter mutation signal is generated; The parameter mutation signal is used to trigger fault diagnosis, wherein the fault diagnosis includes: Analyze the equipment operation data corresponding to the parameter mutation signal and check whether there are abnormal fluctuations in other process parameters during the same period; Based on the analysis results, the type of equipment degradation is correlated to determine the root cause; Following step S104, step S106 is also included: analyzing the statistical characteristics of the predicted residuals and matching them with a failure mode library, wherein the analysis includes: Calculate the residual between the predicted value and the actual value, and calculate the statistical characteristics of the residual, including at least one of variance, mean, skewness or kurtosis; The statistical features are compared with the mapping rules in the preset fault mode library, which is trained based on historical fault data, and the mapping rules include the correspondence between fault types and residual statistical features. Based on the matching results, determine the fault type and alarm level.

2. The adaptive ARIMA modeling method as described in claim 1, characterized in that, Step S105 further includes: if in three consecutive calculations, Δd changes in the same direction, and each Δd is greater than k×σ delta If k is a preset coefficient, then it is determined to be a parameter mutation, and a parameter mutation signal is generated.

3. The adaptive ARIMA modeling method as described in claim 1, characterized in that, In step S103, fitting the ARIMA model includes: The stationarity of the process parameter data within the training window is tested using the extended Dickey-Fuller test. If the p-value is greater than 0.05, the difference operation is performed until the sequence is stationary. The difference order is denoted as d. For stationary sequences, the optimal autoregression order p and moving average order q can be determined through autocorrelation function and partial autocorrelation function analysis, or by using the auto_arima algorithm for grid search.

4. The adaptive ARIMA modeling method as described in claim 1, characterized in that, The equipment status warning signal includes a warning level and a severity level. After the equipment status warning signal is generated, it is sent to the equipment maintenance system or engineer terminal through an interface.

5. An adaptive ARIMA modeling system for predictive maintenance of semiconductor manufacturing equipment, characterized in that, include: The data acquisition module is configured to acquire process parameter data of semiconductor manufacturing equipment in real time, wherein the process parameter data includes at least one of etching rate, temperature or pressure. A window adjustment module is configured to dynamically adjust the training window size of the ARIMA model based on the volatility of the process parameter data. The window adjustment module includes: The volatility calculation unit is configured to calculate the rolling standard deviation σ of recent data. roll The rolling standard deviation σ roll It is the standard deviation of the most recent m data points, where m is a preset positive integer; The threshold comparison unit is configured to compare σ roll With upper limit threshold Th high and lower limit threshold Th low Comparison, where Th high =k1×σ process ,Th low =k2×σ process k1 and k2 are preset coefficients greater than 0, σ process The historical stability standard deviation of the process parameters during the stable operation phase of the equipment; Window control unit, configured to operate based on rolling standard deviation σ roll With upper limit threshold Th high and lower limit threshold Th low Adjusting the training window size based on the comparison results includes: If σ roll >Th high Then increase the training window size to min(W) old ×a,W max ), where W old Where W is the current window size, 'a' is a preset coefficient greater than 1, and W is the current window size. max Set the maximum window size to the preset value; If σ roll <Th low Then reduce the training window size to max(W) old ×b,W min ), where b is a preset coefficient less than 1, W min Set the minimum window size; If Th low ≤σ roll ≤Th high If so, the training window size remains unchanged; The model fitting module is configured to use the adjusted process parameter data within the training window to fit an ARIMA model, wherein the ARIMA model includes an autoregressive order p, a differencing order d, and a moving average order q; and to monitor changes in the ARIMA model parameters, including: Calculate the difference order d of the currently fitted ARIMA model. current The difference order d with the previous fitted model previous The difference Δd = |d current -d previous |; Compare Δd with the control upper limit UCL, where UCL = c × σ delta c is a preset coefficient greater than 0, σ delta The standard deviation of the historical d-value change sequence during the equipment health and stability phase is defined by the time period in historical data when the fluctuation of process parameters is less than a preset threshold. If Δd > UCL, it is determined to be a parameter mutation, and a parameter mutation signal is generated; The parameter mutation signal is used to trigger the operation of the fault diagnosis module, wherein the operation of the fault diagnosis module includes: Analyze the equipment operation data corresponding to the parameter mutation signal and check whether there are abnormal fluctuations in other process parameters during the same period; Based on the analysis results, the type of equipment degradation is correlated to determine the root cause; The prediction and early warning module is configured to use the fitted ARIMA model to predict the next process parameter data and generate equipment status early warning signals based on the prediction results. Generating an equipment status warning signal includes: using a fitted ARIMA model to predict the process parameters for the next process cycle and obtaining predicted values; after the actual completion of the process cycle, obtaining the actual values ​​of the process parameters and calculating the residual between the predicted and actual values; comparing this residual with a pre-set threshold, which is set based on the standard deviation of historical prediction residuals; if the residual exceeds the threshold, generating an equipment status warning signal. The fault matching module is configured to analyze the statistical characteristics of the predicted residuals and match them with a fault mode library. The analysis includes: Calculate the residual between the predicted value and the actual value, and calculate the statistical characteristics of the residual, including at least one of variance, mean, skewness or kurtosis; The statistical features are compared with the mapping rules in the preset fault mode library, which is trained based on historical fault data, and the mapping rules include the correspondence between fault types and residual statistical features. Based on the matching results, determine the fault type and alarm level.

6. An electronic device, characterized in that, include: A memory and a processor, wherein the processor and the memory are connected; The memory is used to store programs; The processor invokes a program stored in the memory to execute the method as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, It stores a computer program, which is executed by a computer to perform the method as described in any one of claims 1-4.

Images