Semiconductor factory affair abnormity attribution method, system and equipment based on machine learning

Abstract

The invention discloses a semiconductor factory affair anomaly attribution method, system and device based on machine learning, and the method comprises the steps: collecting the historical data of an abnormal event and a normal event, screening the related features having the influence on an abnormal result through correlation calculation, and forming a training data set containing a dichotomy label; based on the training data set, constructing a dichotomy tree model and training the dichotomy tree model; based on the dichotomy tree model, outputting a feature importance sequence, and calculating an upper whisker value and a lower whisker value of each feature data in a training data set; when an abnormal event occurs, obtaining a value of a current feature which is the same as the screened feature; and sequentially calculating whether each current characteristic value exceeds the calculated upper required value or the calculated lower required value according to the importance sequence of the output characteristics, if so, marking the characteristics as abnormal attribution characteristics, and outputting an attribution result containing the abnormal attribution characteristics and the sequence of the abnormal attribution characteristics. According to the invention, the consumption of computing resources can be reduced, and sortable decision support can be provided.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of semiconductor manufacturing, and particularly relates to a semiconductor factory operation abnormality attribution method, system and device based on machine learning. BACKGROUND

[0002] With the continuous evolution of semiconductor manufacturing processes, the complexity of semiconductor factory operation systems (covering energy supply, process equipment, environmental control, etc. subsystems) increases exponentially, and the stability of energy supply (such as natural gas, electricity, and special gases) and the running state of key equipment (such as reaction cavities, vacuum pumps, and heating modules) directly determine the yield and production safety of semiconductor wafers. For example, natural gas is the core energy for reaction cavity heating and tail gas incineration system, and its consumption anomaly may cause process temperature fluctuations, which in turn may cause wafer defects; small deviations in equipment operating parameters (such as reaction cavity temperature and vacuum degree) may also cause batch product scrap. Therefore, the semiconductor factory operation system has high requirements for the rapid identification and accurate attribution of energy supply anomalies and equipment anomalies.

[0003] Currently, in the semiconductor factory operation system, correlation analysis between points or setting specification line thresholds are usually used to find the abnormal causes of energy supply anomalies (such as natural gas consumption anomalies) or equipment operation anomalies, such as calculating the correlation coefficient of different sensor data points or judging whether it is out of limit based on a preset threshold; however, these methods have the problems of large amount of calculation and low efficiency, because correlation analysis needs to process a large amount of real-time data, resulting in high resource consumption, and the interpretation based on fixed thresholds is single and cannot adapt to the feature distribution of different events. In addition, the existing technology fails to sort the attribution features by importance, making engineers lack priority basis when making decisions, reducing the accuracy and response speed of abnormal processing. Therefore, there is a main technical problem: how to efficiently and accurately identify and sort the attribution features of abnormal events to reduce the consumption of computing resources and provide sortable decision support, thereby improving the accuracy and practicality of abnormal analysis. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a semiconductor factory operation abnormality attribution method, system, device and medium based on machine learning, which can efficiently and accurately identify and sort the attribution features of abnormal events, reduce the consumption of computing resources, and provide sortable decision support.

[0005] The first aspect of the present application discloses a semiconductor factory operation abnormality attribution method based on machine learning, comprising the following steps: S1. Collecting historical data of abnormal events and normal events, screening relevant features having influence on abnormal results through correlation calculation, forming a training data set containing binary classification labels, wherein the binary classification labels are used to identify abnormal events or normal events; S2. Based on the training data set, constructing and training a binary classification tree model until the model reaches a predetermined accuracy requirement; S3. Based on the binary classification tree model, outputting feature importance ranking, and calculating the upper whisker and lower whisker of the box plot based on each feature data in the training data set, wherein the upper whisker and lower whisker are used to define the normal value interval of the feature; S4. When an abnormal event occurs, obtaining the value of the current feature which is the same as the feature screened out in step S1. S5. According to the feature importance ranking output in step S3, sequentially calculating whether each current feature value exceeds the upper whisker or lower whisker calculated in step S3, if so, marking the feature as an abnormal attribution feature, and outputting the attribution result containing the abnormal attribution feature and its ranking.

[0006] In the above method, in step S1, the correlation calculation is Pearson correlation coefficient method or maximum information coefficient method, which is used to screen relevant features having influence on abnormal results.

[0007] In the above method, in step S2, the binary classification tree model is a random forest model or a LightGBM model, wherein the binary classification tree model training process includes optimizing tree depth and complexity.

[0008] In the above method, the optimization of tree depth and complexity includes the following specific steps: S2.1, in the training process, iteratively training the binary classification tree model and verifying the model performance, wherein the performance verification is based on the validation data set to calculate the classification accuracy rate index; S2.2, when the classification accuracy rate does not reach the predetermined accuracy requirement, adjusting the tree depth parameter and the leaf node number parameter, and repeating step S2.1 until the classification accuracy rate meets the accuracy requirement.

[0009] In the above method, in step S3, the feature importance ranking is quantitatively output through node split gain index or coverage sample proportion index.

[0010] In the above method, in step S3, the upper whisker and lower whisker of the box plot are calculated based on the interquartile range algorithm, including: calculating the first quartile Q1 and the third quartile Q3 of the feature data, obtaining the interquartile range IQR = Q3-Q1, setting the upper whisker as Q3+1.5×IQR, and setting the lower whisker as Q1-1.5×IQR.

[0011] The method, in step S5, the process of sequentially calculating whether each current feature value exceeds the upper whisker value or the lower whisker value, comprises: preferentially detecting a feature with a high feature importance ranking, and if the current feature value of the feature exceeds the upper whisker value or the lower whisker value, immediately marking the feature as an abnormal attribution feature and stopping subsequent feature detection.

[0012] The second aspect of the present application discloses a semiconductor factory abnormal attribution system based on machine learning, comprising a data acquisition module, a model training module, a feature threshold generation module, a feature screening module and an abnormal attribution module. The data acquisition module is configured to collect historical data of abnormal events and normal events, screen relevant features having an impact on abnormal results through correlation calculation, and form a training data set containing binary classification labels, wherein the binary classification labels are used to identify abnormal events or normal events. The model training module is configured to construct a binary classification tree model and train the model according to the training data set until the model reaches a predetermined accuracy requirement. The feature threshold generation module is configured to output a feature importance ranking according to the binary classification tree model, and calculate upper whisker values and lower whisker values of a box plot based on each feature data in the training data set, wherein the upper whisker values and the lower whisker values are used to define a normal numerical interval of the feature. The feature screening module is configured to acquire values of current features same as the features screened in the data acquisition module when an abnormal event occurs. The abnormal attribution module is configured to sequentially calculate whether each current feature value exceeds the upper whisker value or the lower whisker value calculated by the feature threshold generation module according to the feature importance ranking output by the feature threshold generation module, and if yes, mark the feature as an abnormal attribution feature and output an attribution result containing the abnormal attribution feature and a ranking thereof.

[0013] The third aspect of the present application discloses an electronic device, comprising a memory and a processor, wherein the processor and the memory are connected. The memory is configured to store a program. The processor invokes the program stored in the memory to execute the method provided in the first aspect of the present application and / or any possible embodiment in combination with the first aspect of the present application.

[0014] The fourth aspect of the present application discloses a computer readable storage medium having a computer program stored thereon, wherein the computer program is run by a computer to execute the method provided in the first aspect of the present application and / or any possible embodiment in combination with the first aspect of the present application.

[0015] Compared with the prior art, the present application has the following advantages: the present application effectively solves the problems of large amount of calculation, low efficiency and lack of attribution ranking in the prior art. Specifically, step S2 constructs and trains a binary classification tree model (such as a decision tree or a random forest), which learns the historical data pattern adaptively through a machine learning algorithm, replaces the correlation analysis or fixed threshold-based calculation-intensive process in the prior art, significantly reduces the calculation resource consumption of real-time processing, and improves the efficiency of anomaly detection; at the same time, step S3 outputs the feature importance ranking, and calculates the upper whisker and lower whisker of the box plot based on the training data set, which defines an objective normal value interval for each feature, which solves the defects of single explanation and inability to provide priority basis in the prior art, and makes the attribution analysis more quantitative and sortable; further, step S5 checks whether the current feature value exceeds the upper whisker or lower whisker according to the feature importance ranking when an abnormal event occurs, and outputs the ranked attribution result, which ensures the accuracy (through box plot statistics) and efficiency (through priority-driven inspection) of attribution, thereby providing engineers with sortable decision support and improving the practicality and accuracy of anomaly processing.

[0016] The technical solutions of the present application will be further described in detail below with the aid of drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The method flowchart of Example 1 of the present application.

[0018] Figure 2 The system module diagram of Example 2 of the present application. DETAILED DESCRIPTION

[0019] Example 1 The specific implementation of the present application will be described below with the example of the attribution analysis of the "abnormally high natural gas consumption" event in a factory.

[0020] As shown in Figure 1 A semiconductor factory anomaly attribution method based on machine learning, comprising the following steps: S1. Collecting historical data of abnormal events and normal events, screening relevant features that have an impact on abnormal results through correlation calculation, and forming a training data set containing binary classification labels, wherein the binary classification labels are used to identify abnormal events or normal events; In practice, the plant system collects data of all events with unusually high natural gas consumption in the past year (labeled as "abnormal"), as well as a large number of events with normal consumption (labeled as "normal"). For each event, collect feature data that may be relevant, such as: air compressor operating power, cooling water temperature, production of a specific process equipment A, ambient temperature, boiler combustion efficiency, etc. Calculate the correlation of each feature with the "abnormal / normal" label using the "Pearson correlation coefficient". Suppose it is found that "air compressor operating power", "cooling water temperature" and "process equipment A production" are the three features with the highest correlation with the abnormal label (absolute value exceeds a preset threshold, such as 0.5), then they are selected as relevant features, together with the "abnormal / normal" label, to form the final data set for training. The data set format is: each sample contains (air compressor power value, cooling water temperature value, equipment A production value, label).

[0021] S2. Based on the training data set, build and train a binary classification tree model until the model meets the predetermined accuracy requirement; In practice, the above screened data set is used. "Random Forest" is selected as the binary classification tree model. The data set is divided into a training set (70%) and a validation set (30%). The initial setting of the maximum depth of the tree (max_depth) is 5, and the number of leaf nodes of each tree (max_leaf_nodes) is limited to 10. Train the model with the training set, then calculate the classification accuracy on the validation set (e.g. the number of correctly predicted samples / total number of samples). Suppose the initial accuracy is 85%, but the predetermined requirement is 90%. Then adjust the parameters: increase max_depth to 7, and increase max_leaf_nodes to 15. Train and validate again, this time the accuracy reaches 91%, meeting the requirement, and stop training. This process ensures that the model can effectively distinguish between normal and abnormal consumption patterns.

[0022] S3. Based on the binary classification tree model, output a feature importance ranking, and based on each feature data in the training data set, calculate the upper whisker and lower whisker of the box plot, wherein the upper whisker and lower whisker are used to define the normal value range of the feature; In practice, the feature importance based on "node split gain" is extracted from the trained random forest model. Suppose the importance ranking is: air compressor operating power > cooling water temperature > process equipment A production. This means that when the model judges abnormalities, "air compressor operating power" plays the largest role.

[0023] For each feature, the upper and lower whisker values are calculated on the boxplot of the "normal" event samples in the training dataset (or all samples, but in practice, it is common to base on normal samples to define the "normal range"). For example, for the "air compressor running power" feature: find all the power values of normal samples, the first quartile (Q1) is 300 kW, and the third quartile (Q3) is 350 kW, then IQR = 350 - 300 = 50 kW. The upper whisker = Q3 + 1.5 x IQR = 350 + 75 = 425 kW; the lower whisker = Q1 - 1.5 x IQR = 300 - 75 = 225 kW. This means that, according to historical normal data, the "normal range" of this feature value is 225 kW to 425 kW. Similarly, calculate the normal range of "cooling water temperature" (e.g. 20°C to 35°C) and the normal range of "process equipment A output" (e.g. 1000 to 1500 pieces), and save these thresholds.

[0024] S4. When an abnormal event occurs, obtain the current values of the same features as those screened in step S1; In implementation, for example, on a certain day, the factory system monitors that the natural gas consumption suddenly abnormally high. The system immediately obtains the values of the "air compressor running power", "cooling water temperature", "process equipment A output" at the current time.

[0025] S5. According to the feature importance ranking output in step S3, in turn, calculate whether each current feature value exceeds the upper or lower whisker value calculated in step S3, if so, mark the feature as an abnormal attribution feature, and output the attribution result containing the abnormal attribution feature and its ranking.

[0026] In implementation, according to the importance ranking obtained in S3 (air compressor power > cooling water temperature > equipment A output), first check the current "air compressor running power" value (assuming 450 kW). Compare it with the normal range (225 kW - 425 kW) calculated and saved for this feature in S3. 450 kW > 425 kW, exceeding the upper whisker, so the system immediately marks "air compressor running power" as one of the "abnormal attribution features" causing the abnormal high natural gas consumption, and outputs the result: "abnormal attribution feature: air compressor running power (ranking 1)". Then check "cooling water temperature" (assuming 22°C, within the range of 20°C-35°C, normal) and "process equipment A output" (assuming 1300 pieces, within the range of 1000-1500 pieces, normal). The final attribution result is that "air compressor running power" is identified as the cause.

[0027] By training tree model based on historical data to obtain feature importance ranking, and dynamically calculating the box plot threshold of each feature based on historical normal data, the key features causing the anomaly (attribution) can be accurately located when detecting new abnormal events, and the results can be ranked by importance, providing more explicit and valuable decision basis for engineers (such as checking air compressor first), which is significantly better than the method relying on fixed threshold or simple correlation analysis.

[0028] In this embodiment, in step S1, the correlation calculation is Pearson correlation coefficient method or maximum information coefficient method, which is used to screen the relevant features that have an impact on abnormal results.

[0029] In implementation, in the case of natural gas consumption, the system can select to use "Pearson correlation coefficient" to calculate the linear correlation between each candidate feature (such as air compressor power, cooling water temperature, etc.) and "abnormal / normal" label, so as to screen the features with high correlation. Pearson coefficient calculation is simple and commonly used, and is suitable for linear relationship; maximum information coefficient can capture more complex nonlinear relationship.

[0030] By specifying the specific technical means of correlation calculation, the feature screening process is more operable and repeatable, ensuring that the screened features indeed have a significant statistical impact on abnormal results.

[0031] In this embodiment, in step S2, the binary classification tree model is a random forest model or a LightGBM model, wherein the binary classification tree model training process includes optimizing tree depth and complexity.

[0032] It should be noted that the two efficient and widely used ensemble tree models naturally have the ability to evaluate feature importance, and are very suitable for the present scheme.

[0033] In this embodiment, the optimization of tree depth and complexity includes the following specific steps: S2.1, in the training process, the binary classification tree model is iteratively trained and the model performance is verified, wherein the performance verification is based on the calculation of classification accuracy index of the verification data set; S2.2, when the classification accuracy does not meet the predetermined accuracy requirement, adjust the tree depth parameter and the leaf node number parameter, and repeat step S2.1 until the classification accuracy meets the accuracy requirement.

[0034] In implementation, in the case of natural gas consumption, when using the random forest model for training: 1. In the training process, after training each model (or one iteration), the classification accuracy (such as the ratio of correctly predicting normal / abnormal) of the model is calculated using the reserved verification data set.

[0035] 2. Assuming the accuracy of the first training is 85%, which is lower than the required 90%, the engineer (or automatic tuning program) will adjust the model parameters: for example, increase the max_depth of the tree from 5 to 7, and increase the max_leaf_nodes from 10 to 15 (for LightGBM, you may adjust the num_leaves parameter).

[0036] 3. Retrain the model using the new parameters.

[0037] 4. Evaluate again on the validation set, assuming the accuracy reaches 91%, which meets the requirement, then the optimization stops and this model is used. If not, continue to adjust the parameters and repeat the training and validation.

[0038] In practice, in the natural gas consumption case, when training the LightGBM model, first set the initial parameters (such as learning rate 0.1, leaf node number 31) and divide 70% of the data as the training set and 30% as the validation set. Then through iterative optimization: adjust the tree depth, leaf node number and regularization parameter (such as increase the leaf node number to 70, enable feature sampling), and continuously train based on the validation set accuracy (target 90%) until the accuracy is improved to 91%. This process uses the built-in gradient sampling and feature bundling technology of LightGBM to achieve efficient training.

[0039] The optimized LightGBM model achieves high-precision classification (91% accuracy), balancing complexity and generalization ability, providing a reliable foundation for subsequent feature importance ranking and anomaly point detection, ensuring faster and more accurate anomaly attribution.

[0040] By providing a specific technical path for model optimization, through iterative adjustment of key structural parameters (depth, leaf number / node number) and evaluation based on validation set performance, the final model obtained has high accuracy and generalization ability, which is the basis for subsequent reliable feature importance analysis and anomaly attribution.

[0041] In this embodiment, in step S3, the feature importance ranking is quantitatively output by the node split gain indicator or the covered sample proportion indicator.

[0042] In practice, in the trained random forest model, for the "air compressor running power" feature, the system calculates the total sum of impurity reduction (such as Gini coefficient) when it is split at all nodes of all trees (i.e. node split gain). The larger the total sum, the greater the contribution of the feature in distinguishing normal and abnormal, and the higher the importance. Alternatively, calculate the proportion of the number of training samples covered when the feature is used to split nodes. The higher the proportion, the higher the importance. Finally, the features are ranked according to the size of the value calculated by the selected indicator.

[0043] By explicitly specifying the basis for calculating the importance ranking of features, these indicators are inherent, computable metrics within the tree model, making the ranking results objective and interpretable.

[0044] In this embodiment, in step S3, the upper whisker value and the lower whisker value of the box plot are calculated based on the interquartile range algorithm, including: calculating the first quartile Q1 and the third quartile Q3 of the feature data, obtaining the interquartile range IQR = Q3 - Q1, setting the upper whisker value as Q3 + 1.5 * IQR, and setting the lower whisker value as Q1 - 1.5 * IQR.

[0045] In implementation, for example, when calculating the normal range of the "air compressor running power" feature: 1. Sort the power values of all "normal" event samples in the training set from small to large.

[0046] 2. Find the value ranked at the 25th position, i.e., the first quartile Q1 (e.g., 300 kW).

[0047] 3. Find the value ranked at the 75th position, i.e., the third quartile Q3 (e.g., 350 kW).

[0048] 4. Calculate the interquartile range IQR = Q3 - Q1 = 50 kW.

[0049] 5. Calculate the upper whisker value = Q3 + 1.5 * IQR = 350 + 75 = 425 kW.

[0050] 6. Calculate the lower whisker value = Q1 - 1.5 * IQR = 300 - 75 = 225 kW.

[0051] This range represents the normal distribution boundary of the feature value based on historical normal data, and values outside this boundary are considered abnormal points.

[0052] By providing a specific statistical method for calculating dynamic thresholds, the method is not sensitive to data distribution skewness, and can automatically adapt to the actual distribution of each feature based on historical normal data to define a reasonable "normal" interval range, which is more scientific and flexible than fixed thresholds.

[0053] In this embodiment, in step S5, the process of sequentially calculating whether each current feature value exceeds the upper whisker value or the lower whisker value includes preferentially detecting features with high importance ranking, and if the current feature value of a feature exceeds the upper whisker value or the lower whisker value, the feature is immediately marked as an abnormal attribution feature and subsequent feature detection is stopped.

[0054] In implementation, the detection process is optimized: the system strictly checks features according to their importance ranking obtained from S3 (e.g., air compressor power > cooling water temperature > equipment A output). If a feature's current value exceeds its box plot threshold, it is immediately marked as an anomalous attribution feature, and the system stops checking subsequent features. For example, in a natural gas consumption anomaly event, the system first checks the most important "air compressor operating power" current value (450 kW). Finding it greater than the previous value (425 kW), it immediately marks it as an "anomaly attribution feature (ranked 1)," and stops checking the "cooling water temperature" and "equipment A output" current values ​​for anomalies. The attribution result only includes "air compressor operating power."

[0055] The above method significantly improves the efficiency of online detection by leveraging the core value of feature importance ranking: the most important feature is most likely the root cause. Once the most important anomalous feature is found, results can be quickly provided without checking secondary features, saving computational resources and accelerating decision response. This is highly practical in real-world industrial scenarios. However, it also means that if the most important feature is normal while the second most important feature is anomalous, the latter will not be detected (this is a trade-off of the strategy).

[0056] Example 2 like Figure 2 As shown, a semiconductor plant anomaly attribution system based on machine learning includes a data acquisition module, a model training module, a feature threshold generation module, a feature filtering module, and an anomaly attribution module. The data acquisition module is used to collect historical data of abnormal and normal events, filter relevant features that have an impact on abnormal results through correlation calculation, and form a training dataset containing binary classification labels, wherein the binary classification labels are used to identify abnormal or normal events. The model training module is used to construct a binary classification tree model based on the training dataset and train it until the model reaches the predetermined accuracy requirement. The feature threshold generation module is used to output the feature importance ranking according to the binary classification tree model, and to calculate the upper and lower whisker values ​​of the box plot based on each feature data in the training dataset, wherein the upper and lower whisker values ​​are used to define the normal numerical range of the feature. The feature filtering module is used to obtain the value of the current feature that is the same as the feature filtered in the data acquisition module when an abnormal event occurs; The anomaly attribution module is used to calculate whether each current feature value exceeds the upper or lower threshold value calculated by the feature threshold generation module based on the feature importance ranking output by the feature threshold generation module. If so, the feature is marked as an anomaly attribution feature, and the attribution result containing the anomaly attribution feature and its ranking is output.

[0057] The embodiment provides a semiconductor factory operation abnormality attribution system based on machine learning, and the system has the same implementation principle and technical effects as the method embodiment in the embodiment 1. For brief description, the system embodiment part does not mention the corresponding content in the embodiment 1.

[0058] Embodiment 3 A computer readable storage medium, which stores a computer program, the computer program is run by a computer to execute the semiconductor factory operation abnormality attribution method based on machine learning in the embodiment 1. Embodiment 4 An electronic device, comprising a memory and a processor, the processor and the memory are connected; The memory is used for storing a program; The processor calls the program stored in the memory to execute the semiconductor factory operation abnormality attribution method based on machine learning in the embodiment 1. It should be noted that the electronic device can be, but is not limited to, a personal computer (PC), a tablet computer, a mobile internet device (MID) and the like. It should be noted that the processor, the memory and other components that can appear in the electronic device are directly or indirectly electrically connected to each other to realize the transmission or interaction of data. For example, the processor, the memory and other components that can appear can be electrically connected to each other through one or more communication buses or signal lines. It should be noted that each embodiment in the specification adopts a progressive manner for description, and each embodiment focuses on the difference from other embodiments. The same and similar parts of each embodiment can be referred to each other. It should be understood that all the systems and methods disclosed in the embodiments of the present application can be implemented by other means. The system embodiments described above are only illustrative, for example, the flowcharts and block diagrams in the drawings show the possible implementation architecture, function and operation of the system, method and computer program product according to the embodiments of the present application. In this regard, each block in the flowchart or block diagram can represent a module, a program segment or a part of code, which includes one or more executable instructions for implementing the specified logic function. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different orders from those described in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions. In addition, the functional modules in the embodiments of the present 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. If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a notebook computer, a server, a mobile phone, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0059] The above description is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A machine learning-based attribution method for semiconductor plant anomalies, characterized in that, Includes the following steps: S1. Collect historical data of abnormal and normal events, filter relevant features that have an impact on abnormal results through correlation calculation, and form a training dataset containing binary classification labels, wherein the binary classification labels are used to identify abnormal or normal events; S2. Based on the training dataset, construct a binary classification tree model and train it until the model reaches the predetermined accuracy requirement; S3. Based on the binary classification tree model, output the feature importance ranking, and based on each feature data in the training dataset, calculate the upper and lower whisker values ​​of the box plot, wherein the upper and lower whisker values ​​are used to define the normal numerical range of the feature; S4. When an abnormal event occurs, obtain the value of the current feature that is the same as the feature selected in step S1; S5. Based on the feature importance ranking output in step S3, calculate in turn whether each current feature value exceeds the upper or lower threshold value calculated in step S3. If so, mark the feature as an anomalous attribution feature and output the attribution result containing the anomalous attribution feature and its ranking.

2. The method according to claim 1, characterized in that, In step S1, the correlation is calculated using the Pearson correlation coefficient method or the maximum information coefficient method, and is used to screen relevant features that have an impact on abnormal results.

3. The method according to claim 1, characterized in that, In step S2, the binary classification tree model is a random forest model or a LightGBM model, wherein the training process of the binary classification tree model includes optimizing the tree depth and complexity.

4. The method according to claim 3, characterized in that, The optimization of tree depth and complexity includes the following specific steps: S2.1 During the training process, the binary classification tree model is iteratively trained and its performance is verified. The performance verification is based on the classification accuracy index calculated on the verification dataset. S2.2 When the classification accuracy does not meet the predetermined accuracy requirement, adjust the tree depth parameter and the number of leaf nodes parameter, and repeat step S2.1 until the classification accuracy meets the accuracy requirement.

5. The method according to claim 1, characterized in that, In step S3, the feature importance ranking is quantified and output using either a node split gain index or a coverage sample ratio index.

6. The method according to claim 1, characterized in that, In step S3, the upper and lower whisker values ​​of the box plot are calculated based on the interquartile range algorithm, including: calculating the first quartile Q1 and the third quartile Q3 of the feature data to obtain the interquartile range IQR = Q3 - Q1, setting the upper whisker value as Q3 + 1.5 × IQR, and setting the lower whisker value as Q1 - 1.5 × IQR.

7. The method according to claim 1, characterized in that, In step S5, the process of sequentially calculating whether each current feature value exceeds the upper or lower beard value includes prioritizing the detection of features with higher feature importance ranking. If the current feature value exceeds the upper or lower beard value, it is immediately marked as an abnormal attribution feature and subsequent feature detection is stopped.

8. A semiconductor plant anomaly attribution system based on machine learning, characterized in that, It includes a data acquisition module, a model training module, a feature threshold generation module, a feature selection module, and an anomaly attribution module; The data acquisition module is used to collect historical data of abnormal and normal events, filter relevant features that have an impact on abnormal results through correlation calculation, and form a training dataset containing binary classification labels, wherein the binary classification labels are used to identify abnormal or normal events. The model training module is used to construct a binary classification tree model based on the training dataset and train it until the model reaches the predetermined accuracy requirement. The feature threshold generation module is used to output the feature importance ranking according to the binary classification tree model, and to calculate the upper and lower whisker values ​​of the box plot based on each feature data in the training dataset, wherein the upper and lower whisker values ​​are used to define the normal numerical range of the feature. The feature filtering module is used to obtain the value of the current feature that is the same as the feature filtered in the data acquisition module when an abnormal event occurs; The anomaly attribution module is used to calculate whether each current feature value exceeds the upper or lower threshold value calculated by the feature threshold generation module based on the feature importance ranking output by the feature threshold generation module. If so, the feature is marked as an anomaly attribution feature, and the attribution result containing the anomaly attribution feature and its ranking is output.

9. 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 1-7.

10. 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 1-7.

Images