Electronic grade anhydrous hydrogen fluoride arsenic removal data analysis system based on decision tree model

Abstract

The application discloses an electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model, belongs to the technical field of decision trees, and comprises a variable sensitive interval hit determination module, a fluctuation judgment processing branch path determination module and a disturbance processing effect analysis module. Through integration of sensitive region identification, disturbance sample construction, prediction fluctuation scoring, model self-adaptive adjustment and prediction smoothing mechanism, the application can identify and relieve the prediction jump problem caused by input disturbance in the arsenic removal process, and improve the stability and precision of the model in the critical interval of the key variable. The system has the ability to realize online adjustment without retraining the main model, has strong adaptability and high calculation efficiency, is suitable for deployment in an industrial-grade scene to perform high-precision and high-robustness real-time prediction and auxiliary decision on the arsenic removal effect, and enhances the data support and intelligent level of the arsenic removal process control.

Description

Technical Field

[0001] This invention relates to the field of decision tree technology, and in particular to a data analysis system for arsenic removal using electronic-grade anhydrous hydrogen fluoride based on a decision tree model. Background Technology

[0002] In the production of electronic-grade anhydrous hydrogen fluoride, arsenic is a critical impurity that requires strict control, with residual levels often needing to be as low as ppb or even lower. Since the removal efficiency of arsenic is influenced by multiple process parameters, including raw material quality, oxidant type and molar ratio, reaction temperature, reaction residence time, crystallization conditions, and separation methods, the variables exhibit complex, nonlinear, multi-coupled, and high-dimensional relationships. Traditional linear regression or univariate models are insufficient to effectively characterize these interaction mechanisms. Therefore, to achieve accurate modeling and predictive control of residual arsenic concentration, decision tree models (such as CART, GBDT, or XGBoost) with nonlinear modeling capabilities and interpretable structures are used as the core data analysis tool.

[0003] The advantages of decision tree models in industrial process data modeling are: their clear structure, with output being a condition-outcome tree path, facilitating direct mapping to process logic; the fact that their training process does not require data normalization, enabling them to adapt to multi-scale and heterogeneous process inputs; and their support for feature importance analysis, PDP, and SHAP interpretation, which helps in identifying key factors influencing residual arsenic concentration and their direction of action. Furthermore, decision tree models exhibit stability in small-to-medium sample environments, making them suitable for the often limited sample size in electronic-grade chemical production processes.

[0004] For example, Chinese invention patent application CN114169537B discloses a federated learning method and system for vertical XGBoost decision trees. It provides a joint training process and a joint inference process for vertical XGBoost decision trees. During the joint training process, split points are calculated, and during the joint inference process, each node is judged. The information disclosed in the joint training process is the maximum split value of each participant, without directly revealing the feature information of each participant. The security of the joint inference process relies on a threshold homomorphic encryption scheme.

[0005] The above-mentioned technology has at least the following technical problems:

[0006] In the application of decision tree models in electronic-grade HF arsenic removal, a problem arises: because the tree model is based on variable threshold splitting, it suffers from discontinuous prediction results near variable boundaries. Taking the NaF molar ratio as an example, a slight adjustment from 5.0 mol to 5.2 mol may trigger a path switch to another leaf node, causing the predicted As concentration to jump abruptly from 18 ppb to 27 ppb, even with a very slight input change. This input perturbation leading to output jumps directly interferes with production control decisions, especially when the predicted value is close to the acceptable threshold, easily causing misjudgments or missed judgments. Summary of the Invention

[0007] To address the aforementioned technical problems in existing technologies, this invention provides a data analysis system for arsenic removal using an electronic-grade anhydrous hydrogen fluoride based on a decision tree model. The technical solution is as follows:

[0008] The variable sensitivity interval hit determination module is used to collect the set of electronic-grade anhydrous hydrogen fluoride arsenic removal data variables to form arsenic removal data input samples, and to determine the hit of variable sensitivity intervals. If no sensitive interval is hit, the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis results are output through the currently executed decision tree model.

[0009] The fluctuation judgment and processing branch path determination module is used to count the number of variables hit when a sensitive interval is hit, thereby generating a disturbance input sample sequence. This sequence is then input into the currently executing decision tree model to obtain the disturbance prediction output sequence. The predicted fluctuation score is analyzed to determine the fluctuation judgment and processing branch path and to perform targeted processing of the disturbance fluctuation.

[0010] The disturbance handling effect analysis module is used to analyze the disturbance handling effect after the targeted handling of disturbance fluctuations is completed, and outputs the data analysis results of electronic-grade anhydrous hydrogen fluoride arsenic removal after effective disturbance handling.

[0011] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0012] 1. The electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model provided by this invention integrates sensitive area identification, perturbation sample construction, prediction fluctuation scoring, model adaptive adjustment, and prediction smoothing mechanisms. This enables the identification and mitigation of prediction jumps caused by input perturbations in the arsenic removal process, improving the model's stability and accuracy within the critical range of key variables. The system possesses the ability to achieve online adjustment without retraining the main model, exhibiting strong adaptability and high computational efficiency. It is suitable for deployment in industrial settings for high-precision, robust real-time prediction and decision support of arsenic removal effects, enhancing the data support and intelligence level of arsenic removal process control.

[0013] 2. This invention, by determining whether a variable hits a sensitive interval, can identify whether the input sample is in a critical region of the model structure prone to prediction jumps before model prediction. Once a variable hits a sensitive interval, the system will trigger a perturbation analysis and stability detection process, avoiding the output discontinuity problem caused by the split structure of the decision tree model, improving the reliability and robustness of the model near fluctuations in core variables, preventing large prediction deviations caused by slight input perturbations, and enhancing the fault tolerance and engineering usability of the decision tree model in actual process control scenarios, thereby ensuring the stability and reliability of the arsenic removal process prediction results.

[0014] 3. This invention, through disturbance uncertainty type analysis, can further identify the source type of fluctuations within the unstable region of model prediction. Specifically, it determines whether the current prediction uncertainty stems from noise disturbances at the input level or from knowledge blind spots at the model structure level. Differentiated response strategies are adopted based on these different types. For uncertainties caused by input noise, smooth fusion enhances prediction robustness; for uncertainties caused by model blind spots, structural fine-tuning is prioritized. This avoids a one-size-fits-all approach to all anomalies, improving the adaptability, sensitivity, and misjudgment tolerance of the decision tree model, providing a more intelligent and stable prediction guarantee for electronic-grade anhydrous hydrogen fluoride arsenic removal processes. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a structural diagram of the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model provided in an embodiment of the present invention.

[0017] Figure 2 This is a flowchart illustrating the disturbance processing effect analysis in this embodiment.

[0018] Figure 3 This is a schematic diagram of the entire process of arsenic removal data analysis using electronic-grade anhydrous hydrogen fluoride based on a decision tree model, according to an embodiment of the present invention.

[0019] Figure 4 This is a model architecture diagram of the decision tree model involved in the embodiments of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0021] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0022] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0023] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0024] To make the technical problem to be solved, the technical solution and advantages of the present invention clearer, a detailed description will be given below with reference to the accompanying drawings for the purpose of data analysis to predict arsenic content.

[0025] like Figure 1 The diagram shown is a data analysis system structure for arsenic removal using an electronic-grade anhydrous hydrogen fluoride based on a decision tree model. It includes: a variable sensitivity interval hit determination module, a fluctuation judgment and processing branch path determination module, and a disturbance processing effect analysis module.

[0026] like Figure 3The diagram illustrates the entire process of arsenic removal data analysis using an electronic-grade anhydrous hydrogen fluoride based on a decision tree model, according to an embodiment of the present invention. First, arsenic removal data input samples are constructed, and a sensitivity interval hit determination is performed. If the determination result is that the sensitivity interval is not hit, the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis result is directly output. If the determination result is that the sensitivity interval is hit, a perturbation input sample sequence is generated, and the predicted fluctuation score is further analyzed. When the predicted fluctuation score is less than the predicted fluctuation score threshold, the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis result is directly output using the arsenic removal data input sample as input. When the predicted fluctuation score is less than the predicted fluctuation score threshold, the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis result is directly output using the arsenic removal data input sample as input. When the score is greater than or equal to the predicted fluctuation score threshold, the system determines the uncertainty type based on whether the predicted result range is less than the predicted result range threshold. If the predicted result range is less than the predicted result range threshold, it is recorded as input noise uncertainty and output smoothing is performed. If the predicted result range is greater than or equal to the predicted result range threshold, it is recorded as model structure uncertainty and lightweight model structure adjustment is performed. Then, the system enters the disturbance processing effect analysis stage. If the result of the disturbance processing effect analysis is effective disturbance processing, the system outputs the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis result; otherwise, it outputs a warning prompt.

[0027] The variable sensitivity interval hit determination module is used to collect the set of electronic-grade anhydrous hydrogen fluoride arsenic removal data variables to form arsenic removal data input samples, and to determine the hit of variable sensitivity intervals. If no sensitive interval is hit, the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis results are output through the currently executed decision tree model.

[0028] Further, a variable sensitivity interval hit determination is performed. The specific determination process is as follows:

[0029] The core variable data of the electronic-grade anhydrous hydrogen fluoride arsenic removal process are collected at a preset collection cycle to form the arsenic removal data input sample.

[0030] In the electronic-grade anhydrous hydrogen fluoride arsenic removal process, precise control of residual arsenic concentration is crucial for ensuring product purity and meeting the demands of downstream high-end processes. To achieve intelligent modeling and stable operation monitoring of the arsenic removal process, the system needs to collect and analyze multiple core variables with engineering quantifiable characteristics in real time. These variables directly affect oxidation reaction efficiency, arsenic species migration behavior, and the final residual concentration. In this embodiment, these variables specifically include: NaF molar ratio, oxidant addition acceleration rate, reaction temperature, average residence time, HF stock solution purity, and initial arsenic concentration.

[0031] Sodium fluoride (NaF) is a commonly used auxiliary precipitant in arsenic removal processes, and its dosage ratio to the molar amount of HF is defined as the NaF molar ratio. The NaF molar ratio significantly affects the stability and precipitation efficiency of the arsenic-fluorine complexes formed in the system. The molar ratio is generally controlled within the range of 4.5 to 6.5 mol / mol, and even slight changes can cause abrupt changes in process behavior, making it one of the highly sensitive variables in the modeling process.

[0032] The efficiency of arsenic oxidation state conversion largely depends on the oxidant (e.g., or The oxidant dosing rate (mL / min) controls the reaction rate and the rate of heat release, playing a decisive role in arsenic formation and subsequent separation. The dosing rate is generally controlled within the range of 0.5 to 2.0 mL / min; too fast or too slow a rate may trigger nonlinear reaction behavior.

[0033] Reaction temperature (°C) is one of the core parameters controlling the oxidation reactivity and the stability of arsenic-fluorine complexes. Excessively high temperatures may trigger side reactions or increase the solubility of arsenic species, affecting arsenic removal efficiency; while excessively low temperatures reduce the reaction rate. In industrial operation, this variable is typically adjusted within the range of 40 to 80°C and is dynamically regulated by the heating / cooling system.

[0034] The average residence time (in seconds) of the reaction liquid in the main reaction section directly determines the degree of completion of the arsenic oxidation and precipitation reactions, and is an important variable characterizing the reactor kinetics. The average residence time is affected by flow rate, liquid level, and equipment structure, and is typically controlled within the range of 60 to 300 seconds. Together with temperature and oxidant, it constitutes the three factors of reaction kinetics.

[0035] The initial purity of anhydrous HF has a fundamental impact on the migration pathway of arsenic and the subsequent behavior of precipitate formation. Unstable water or impurity components in HF will affect the speciation and reaction pathway of arsenic. The purity (%) of the HF stock solution is obtained through raw material testing and is generally not collected in real time, but is input as a static characteristic variable into the modeling system.

[0036] The initial arsenic concentration in the raw solution before entering the reaction stage determines the system's load level and is a crucial baseline input for model prediction. The initial arsenic concentration (ppb) may originate from factors such as batch variations in raw materials and equipment residues. It is typically measured using online ICP-MS or laboratory samples, requiring high precision, and generally ranges from 20 to 150 ppb.

[0037] It should be noted that the above variables are only the core variables set in this embodiment. In specific embodiments, specific settings can be made according to specific circumstances. This embodiment does not impose any limitations.

[0038] Extract the sensitivity interval matching table of each core variable built into the system.

[0039] In the process of analyzing arsenic removal data using electronic-grade anhydrous hydrogen fluoride based on decision tree models, the structural division of these models relies on variable splitting thresholds. Therefore, even minor perturbations in some input variables can easily trigger jumps in the prediction path, leading to abrupt changes or instability in the model's output. To proactively identify such high-risk input scenarios during system operation and ensure prediction stability and result reliability, the system designs a core variable sensitive interval matching table as a pre-triggered mechanism for perturbation analysis and judgment.

[0040] The core variable sensitivity interval matching table is a predefined structured data table used to record the splitting threshold intervals of variables that are highly correlated with the model structure and sensitive to predictive behavior. The construction of this table typically involves the following steps:

[0041] Model structure analysis involves analyzing the structure of the currently trained decision tree model, extracting the key split points and their statistical frequencies of all split variables, and identifying regions that are prone to triggering path changes.

[0042] Backtracking analysis of prediction fluctuations identifies which input value ranges exhibit significant prediction fluctuations or model switching behavior within historical sample prediction records.

[0043] By superimposing critical business risk values ​​and combining them with the actual control critical points of key variables in the process flow (such as the NaF molar ratio approaching the process alarm value), a safety buffer zone is further set.

[0044] Ultimately, the core variable sensitive interval matching table is organized in the form of key-value pairs or structured dictionaries, with each record including fields such as variable name and sensitive interval range (e.g., [4.8, 5.3]mol).

[0045] The input samples of arsenic removal data are judged based on the sensitive interval matching table of each core variable. If no core variable in the input samples of arsenic removal data falls into its corresponding sensitive interval, the result of the variable sensitive interval hit judgment is recorded as a non-hit sensitive interval.

[0046] If no core variable in the arsenic removal data input sample falls within its corresponding sensitive interval, it means that the key operating parameters of the current electronic-grade anhydrous hydrogen fluoride arsenic removal process (such as NaF molar ratio, oxidant dosing rate, reaction temperature, etc.) are all within a relatively stable and safe normal operating range. In this case, the variable sensitive interval hit determination result is "not hit," indicating that the process operation status has not encountered any abnormal variables that could cause drastic fluctuations in the prediction results or distortion of the model judgment. The system can directly process the arsenic removal data input sample through the currently executed decision tree model and output stable and reliable arsenic removal data analysis results without initiating additional perturbation processing procedures.

[0047] In a specific embodiment, when the sensitive region is not hit, the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis result is output through the currently executed decision tree model. Specifically, the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis result is output through the currently executed decision tree model, using the arsenic removal data input sample as input. The aforementioned currently executed decision tree model refers to the version of the decision tree model actually used to perform prediction calculations on the input data sample at the current time and under the current system configuration. It has the following characteristics: it has been trained, it has been deployed online, and it may have undergone lightweight fine-tuning.

[0048] like Figure 4 The diagram shows the model architecture of the decision tree model involved in this embodiment of the invention. It includes a data input stage, a subtree structure layer, a subtree fusion module, and a final output node. In the data input stage, arsenic removal data input samples are used as input to the decision tree model. Specifically, the arsenic removal data input samples at the top of the diagram represent the core variable data input of the electronic-grade HF arsenic removal process collected by the system. These variables include, but are not limited to, NaF molar ratio, oxidant addition acceleration rate, and reaction temperature, all of which are structured numerical data with clear dimensions, constituting a complete input sample vector. The subtree structure layer includes main model subtree structure 1, main model subtree structure 2, and main model subtree structure 3. The three subtree structures in the middle represent three typical subtrees in the currently executed decision tree ensemble model (such as XGBoost). They constitute the structural basis of the main model. Each tree is modeled based on different training samples or feature subsets, and has independent splitting paths and prediction behaviors. This multi-subtree ensemble structure helps enhance the model's generalization ability and local stability. The subtree fusion module is used to fuse the output information of the subtree structure layer and output a unified predicted value. Fusion methods include, but are not limited to, weighted average, model confidence weighting, and perturbation sample weighting. When input perturbations exist or variable boundary regions are hit, the system can dynamically adjust the weights of each subtree to improve the stability and accuracy of the prediction. The final output node outputs the residual arsenic concentration prediction result. In specific embodiments, the residual arsenic concentration prediction result directly serves downstream quality assessment, parameter optimization suggestions, or process stability control.

[0049] It should be noted that the diagram uses a three-subtree structure as an example, only to illustrate the core mechanism of the invention. In practical applications, the number of subtrees can be flexibly adjusted according to specific engineering needs, data complexity, or computing resource allocation; the invention does not impose any restrictions on this.

[0050] In this embodiment, to achieve high-precision prediction of residual arsenic content during the removal of arsenic by electronic-grade anhydrous hydrogen fluoride, the system constructs a supervised learning model based on historical process operation data, employing an interpretable boosting decision tree structure as the main model (such as regression tree ensemble implemented by XGBoost or LightGBM). The input data used for model training includes six key process variables: NaF molar ratio, oxidant dosing rate, reaction temperature, average residence time, HF stock purity, and initial arsenic concentration. These variables are all numerical, quantifiable, and continuous features. The system first standardizes the original batch data, including missing value completion, outlier removal, and normalization transformation, to ensure the stability and usability of the training data. After feature preparation, a training set of sample pairs (X, y) is constructed, where X is a six-dimensional input vector and y is the corresponding residual arsenic concentration value. During training, a gradient boosting strategy is used to iteratively generate weak regression trees using mean squared error as the loss function, and the overall prediction performance is improved through ensemble analysis.

[0051] To avoid overfitting and enhance generalization ability, a cross-validation mechanism was introduced during training, and grid search optimization was performed on several key hyperparameters (such as maximum tree depth, minimum number of samples per leaf node, minimum split gain, and regularization weights). After training, the system permanently deployed the model to the online prediction process as the current execution decision tree model, supporting subsequent real-time perturbation analysis and prediction stability assessment processes.

[0052] If a core variable in the arsenic removal data falls into its corresponding core variable sensitive interval, the result of the variable sensitive interval hit determination is recorded as the hit sensitive interval, and the core variable is simultaneously recorded as the hit variable.

[0053] The fluctuation judgment and processing branch path determination module is used to count the number of variables hit when a sensitive interval is hit, thereby generating a disturbance input sample sequence. This sequence is then input into the currently executing decision tree model to obtain the disturbance prediction output sequence. The predicted fluctuation score is analyzed to determine the fluctuation judgment and processing branch path and to perform targeted processing of the disturbance fluctuation.

[0054] If a core variable in the arsenic removal data falls within its corresponding sensitive interval, it indicates that at least one key operating parameter (such as the NaF molar ratio, oxidant dosing rate, reaction temperature, etc.) in the current electronic-grade anhydrous hydrogen fluoride arsenic removal process is within a critical range that may cause uncertainty in model predictions. In this case, the variable sensitive interval hit determination result is "hit sensitive interval," and the core variable falling within the sensitive interval will be marked as a hit variable. This situation means that the current parameter state may cause problems such as output jumps and increased dispersion in the decision tree model's prediction of residual arsenic content. Therefore, the system needs to further count the number of hit variables, generate a perturbation input sample sequence, and determine the subsequent processing path by analyzing the prediction fluctuation score to avoid interference from model prediction distortion on process judgment.

[0055] Furthermore, the number of hit variables is counted, thereby generating a perturbed input sample sequence. The specific process is as follows:

[0056] The number of hit variables is counted. If the number of hit variables is 1, a perturbation value is constructed based on the hit variable and the hit variable data, and a perturbation input sample sequence of the hit variable is generated.

[0057] In a specific embodiment, when only one input variable falls within the sensitive interval (e.g., the NaF molar ratio falls within [4.8, 5.3] mol), the goal of perturbation analysis is to examine the impact of small changes in this variable within its neighborhood on the model's prediction results. Since the variable is singular, there is no need to construct a multidimensional space, and there are no interactions between variables. In this case, the central perturbation method (e.g., ±0.1, ±0.2, etc.) can efficiently construct the perturbation input sequence. The advantages of this method are its simplicity, low computational cost, and ability to accurately characterize the fluctuation trend of the prediction curve within the local region of the current variable, directly reflecting the path jump behavior of the decision tree model near the split point.

[0058] If the number of hit variables is greater than 1, the Latin hypercube sampling method is used. For each hit variable, it is divided into K equal-width sub-intervals based on its corresponding sensitive interval. Then, random selection is made in each sub-interval to form the perturbation sample set of the hit variable. The perturbation sample sets of each hit variable are combined, shuffled and rematched to generate a linked perturbation sample sequence, which is recorded as the perturbation input sample sequence.

[0059] It should be added that K represents the number of sub-intervals into which the sensitive interval is divided. In specific implementations, the specific value of K should be dynamically set according to the actual engineering scenario. When the sensitive interval is large, a larger value of K can be selected to refine the perturbation; if the sensitive interval itself is narrow, the value of K should not be too large, so as to avoid the sub-intervals being too narrow and resulting in no significant difference between the perturbation samples.

[0060] When multiple input variables (such as oxidant addition acceleration rate, reaction temperature, average residence time, etc.) simultaneously hit the sensitive region, there are synergistic or coupled relationships between the variables, making it impossible to simply perturb them one by one. If Cartesian product enumeration is still used in this case, the number of samples grows exponentially, making it difficult to meet the response time requirements of the online system; if random sampling is used, the sample uniformity is insufficient, and the coverage is unstable. Therefore, the system chooses to use Latin hypercube sampling as the perturbation sample generation method.

[0061] The Latin hypercube sampling method ensures that the perturbation interval of each variable is representatively sampled by equally dividing the interval within each sensitive variable. Furthermore, it introduces a shuffling and rearrangement mechanism during combination matching to generate a uniformly distributed and information-rich set of perturbation input samples. This method exhibits good coverage and stability in high-dimensional perturbation spaces and can effectively reveal changes in the model response surface morphology under multiple variable perturbations.

[0062] In a specific embodiment, assuming the current situation is a two-dimensional perturbation, and the hit variables are the NaF molar ratio and the reaction temperature, and assuming the perturbation range of the NaF molar ratio is set to [4.8, 5.3] and the perturbation range of the reaction temperature is [80.95] in the core variable sensitivity interval matching table, and K is 5, then the specific process of obtaining the perturbation input sample sequence is as follows:

[0063] A1, divide the NaF molar ratio into five segments: [4.8, 4.9), [4.9, 5.0), [5.0, 5.1), [5.1, 5.2), [5.2, 5.3].

[0064] The reaction temperature was divided into five segments: [80,83), [83,86), [86,89), [89,92), [92,95].

[0065] A2, randomly sample a value in each sub-interval, for example:

[0066] NaF molar ratio sampling: {4.83, 4.97, 5.04, 5.21, 5.12}.

[0067] Reaction temperature sampling: {80.5, 85.2, 88.7, 92.1, 83.3}.

[0068] A3, after shuffling the sample set, combines them to form 5 two-dimensional perturbation inputs:

[0069] X1'=(4.83,88.7), X2'=(4.97,80.5), X3'=(5.04,83.3), X4'=(5.21,85.2), X5'=(5.12,92.1).

[0070] This results in a sample set {X1', X2', X3', X4', X5'}, which is the perturbed input sample sequence, where each X' is a d-dimensional vector (d=2 in this example).

[0071] Furthermore, the perturbation prediction output sequence is obtained, and the specific acquisition process is as follows:

[0072] The system loads the decision tree model currently used for the electronic-grade anhydrous hydrogen fluoride arsenic removal process data, calls the model inference interface, and inputs the perturbation input sample sequence as a batch of data into the currently executing decision tree model.

[0073] The current execution decision tree model predicts the value of the residual arsenic content for each perturbation input and output, representing the predicted value of the perturbation combination. All output results are combined into a prediction sequence, which is denoted as the perturbation prediction output sequence.

[0074] Furthermore, the fluctuation score is analyzed and predicted. The specific process is as follows:

[0075] Based on the perturbation prediction output sequence, the maximum and minimum values ​​of the output prediction results are extracted. The difference between the maximum and minimum values ​​of the prediction results is then processed, i.e., the maximum value of the prediction results is subtracted from the minimum value of the prediction results, to obtain the prediction range. The prediction range is used to measure the upper and lower bound fluctuation range of the prediction results.

[0076] Calculate the standard deviation of the perturbation prediction output sequence, denoted as the prediction result standard deviation. The prediction result standard deviation is used to measure the overall dispersion of the prediction results.

[0077] The perturbation prediction output sequence is fitted to a one-dimensional linear trend line. The root mean square error between the perturbation prediction output sequence and the fitted value is calculated and denoted as the trend residual score of the prediction result. The trend residual score of the prediction result is used to capture nonlinear jumps or trend abrupt changes.

[0078] In the data analysis of arsenic removal using electronic-grade anhydrous hydrogen fluoride, when an input variable hits the sensitive region of the model, the system generates a set of perturbation input samples. These perturbation samples are then sequentially input into the currently executing decision tree model to obtain the corresponding perturbation prediction output sequence. This sequence records the changes in the model output response to input fine-tuning within the input neighborhood, providing numerical evidence for analyzing the model's local stability and response trends.

[0079] To further evaluate the trend stability of the model output, the system treats the perturbation prediction output sequence as a set of discrete numerical points and performs a one-dimensional linear regression fit to generate an optimal linear trend line. This trend line reflects the overall direction of change of the predicted value with the perturbation input, i.e., the system's tendency to linearly respond to the perturbation of this variable. If the model's output is relatively stable within the perturbation interval, its predicted value should closely match the trend line, indicating that the model has a stable response to small changes in this variable. Subsequently, the system calculates the root mean square error between the perturbation prediction value and its corresponding linear fit value and defines it as the prediction result trend residual score. The prediction result trend residual score quantifies the degree of fluctuation of the model output and the degree of deviation from the linear trend, and is the core indicator for evaluating whether the model's predicted trend is smooth within the current input neighborhood. If the prediction result trend residual score is small, it indicates that the perturbation prediction is approximately linear and changes smoothly, and the model has good stability; if the prediction result trend residual score is large, it indicates that the model is sensitive to perturbations and may have structural instability or overfitting problems, requiring further optimization of the model structure or the introduction of an output smoothing mechanism.

[0080] By setting a scaling factor, the prediction result range, prediction result standard deviation, and prediction result trend residual score are scaled down, and the maximum value after scaling down is taken as the prediction fluctuation score.

[0081] In the data analysis workflow for arsenic removal using anhydrous hydrogen fluoride in electronic grade, the core purpose of setting the scaling factor is to normalize three indicators (prediction result range, prediction result standard deviation, and prediction result trend residual score) with different physical meanings and numerical dimensions, allowing them to be compared on the same evaluation scale. This ensures that the composite fluctuation scoring mechanism, which replaces the weighted average with the maximum value, is reasonable and discriminative. The general logic and basis for setting the scaling factor are as follows: The scaling factor set in the system is a set of empirical parameters, corresponding to the three core prediction fluctuation indicators. Its setting method can be summarized into the following three core principles:

[0082] B1, Historical Data-Driven Dimensional Scaling: By statistically analyzing the common value ranges of each type of fluctuation indicator in historical disturbance prediction data, a target normalization interval is set. For example, the predicted range is typically distributed between 2 and 10 ppb; the predicted standard deviation is mostly between 1 and 5 ppb; and the trend residual score is mostly concentrated between 0.5 and 3 ppb. For these value ranges, the system can select a unified target interval (e.g., 0 to 1) as the standardized output interval, and then define a linear scaling coefficient for each type of indicator to ensure that it falls within the target range after normalization. For example, if the target standardized value range is [0,1], and the typical upper limit of the standard deviation in historical samples is 5 ppb, then the standard deviation scaling coefficient can be set to 1 / 5.

[0083] B2. Sensitivity Adjustment Guided by Engineering Tolerance: Different industrial scenarios have varying tolerances for different types of fluctuations. For example, the prediction range represents the significant leap between the upper and lower limits of the prediction; its abrupt changes are more likely to trigger control misjudgments and should be assigned higher sensitivity. Conversely, deviations in trend residuals may indicate nonlinear changes or boundary anomalies within the system, requiring enhanced response in certain demanding processes. Therefore, scaling factors are not merely unit scaling of physical quantities but can also reflect engineering priorities. Larger scaling ratios can be set for indicators with low tolerance, making them more likely to trigger high-risk assessments. These coefficients can be tuned by engineers or adaptively learned through methods such as Bayesian optimization.

[0084] B3, Continuous Calibration of Cross-Batch Stability: The scaling factor is not statically configured but should possess a self-calibration mechanism. The system can periodically evaluate the correlation between the predicted fluctuation score and the actual arsenic removal error, fine-tuning the scaling factor to ensure the final score is most correlated with the actual error. This mechanism can provide dynamic feedback based on the evaluation accuracy within a sliding window, thereby enhancing the robustness of the model during long-term operation.

[0085] In summary, the core value of preset scaling factors lies in unifying fluctuation indicators of different dimensions into a comparable evaluation framework, while retaining the ability to identify key anomaly patterns. Its setting is based on a three-pronged approach: historical value range analysis, engineering sensitivity prioritization, and a dynamic adaptive update mechanism. This is one of the key foundations for ensuring the effectiveness, stability, and practicality of the entire predictive fluctuation identification mechanism.

[0086] In a specific embodiment, the predicted volatility score is represented as follows:

[0087] ,

[0088] Where A is the predicted volatility score, a is the predicted range, b is the predicted standard deviation, c is the predicted trend residual score, a1 is the scaling factor for the predicted range, b1 is the scaling factor for the predicted standard deviation, and c1 is the scaling factor for the predicted trend residual score.

[0089] Furthermore, the fluctuation judgment and processing branch path is determined, and the specific analysis process is as follows:

[0090] Extract the preset predicted volatility score threshold from the database.

[0091] In this embodiment, the predicted volatility score threshold is used to determine whether the current model response volatility is within an acceptable range. The preset process for the predicted volatility score threshold in this embodiment is as follows: the system extracts a large number of input samples located in non-sensitive regions from the historical sample library, calculates the predicted volatility score under its perturbation sample group, and forms a volatility score distribution curve under a stable state. Subsequently, the 90th percentile value or the verified optimal stability lower bound is selected from this distribution as the predicted volatility score threshold.

[0092] It should be noted that the above-described process for setting the predicted fluctuation score threshold is only an example. In specific embodiments, the setting of the predicted fluctuation score threshold can be made in combination with actual working conditions and requirements. This embodiment does not limit it.

[0093] If the predicted fluctuation score is less than the predicted fluctuation score threshold, the fluctuation judgment processing branch path is recorded as directly using the arsenic removal data input sample as input, and outputting the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis results based on the current execution decision tree model.

[0094] If the predicted fluctuation score is less than the predicted fluctuation score threshold, it indicates that the overall fluctuation range of the predicted results obtained by perturbing the input sample sequence (such as the predicted value of residual arsenic content) is small, the dispersion is low, and there is no obvious nonlinear jump or trend change behavior. This shows that the currently executed decision tree model is minimally affected by variable perturbation when processing this arsenic removal data input sample, and the model prediction stability and reliability are high, requiring no adjustment to the prediction results or model structure. Therefore, the fluctuation judgment processing branch path is determined to directly use the original arsenic removal data input sample as input, calculate and output the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis results through the currently executed decision tree model, ensuring both analytical efficiency and result accuracy.

[0095] If the predicted volatility score is greater than or equal to the predicted volatility score threshold, then a disturbance uncertainty type analysis is performed. If the predicted result range is less than the preset predicted result range threshold, then the disturbance uncertainty type is recorded as input noise uncertainty, and the volatility judgment and processing branch path is recorded as output smoothing.

[0096] If the predicted volatility score is greater than or equal to the predicted volatility score threshold, it indicates that the fluctuation in the predicted result caused by the perturbation input sample sequence exceeds the system's allowed stability range, potentially posing a risk of prediction distortion. Further analysis of the specific type of perturbation uncertainty is needed to develop a targeted handling plan. This situation typically means that variable perturbations in the current arsenic removal data input sample have already affected the prediction output of the decision tree model. Directly using the original model output results may lead to errors in production control judgments (such as misjudging whether the residual arsenic content meets the standard). Therefore, it is necessary to initiate subsequent uncertainty type analysis and corresponding processing procedures, rather than directly outputting the results.

[0097] If the predicted range is less than the predicted range threshold, it indicates that the overall fluctuation of the predicted result mainly stems from minor noise interference in the input data (such as random errors in sensor data acquisition or subtle random fluctuations in process parameters), rather than the model's own structure's sensitive response to variable changes. In other words, the disturbance uncertainty type is input noise uncertainty. At this point, the upper and lower bounds of the predicted result are still within a controllable range. There is no need to adjust the model structure; simply executing the output smoothing process can reduce the impact of noise on the final result and ensure the stability of the output result.

[0098] Furthermore, output smoothing is performed, and the specific analysis process is as follows:

[0099] The predicted volatility score is obtained by subtracting the predicted volatility score threshold from the predicted volatility score.

[0100] Use the predicted fluctuation deviation score as the query key to query the smoothing fusion coefficient.

[0101] In this embodiment, the system pre-constructs a configuration table containing the correspondence between predicted fluctuation deviation scores and smoothing fusion coefficients. This table needs to be generated based on a large amount of measured data or simulation results, and clearly defines the smoothing fusion coefficients matched with different predicted fluctuation deviation score intervals (low deviation scores correspond to smaller smoothing fusion coefficients to reduce over-smoothing, and high deviation scores correspond to larger coefficients to enhance the data correction effect). Then, the specific predicted fluctuation deviation score is obtained through analysis. Finally, the calculated predicted fluctuation deviation score is used as a query key value to perform interval matching or precise search in the preset configuration table, so as to quickly obtain the smoothing fusion coefficient that matches the current predicted fluctuation deviation state, ensuring that the smoothing fusion coefficient can accurately respond to the data predicted fluctuation deviation, and improving the accuracy and adaptability of data processing.

[0102] The smoothing fusion coefficient is a scaling factor used to weight and fuse the original prediction results with the mean of the perturbation predictions. Its value ranges from 0 to 1. In specific embodiments, when the smoothing fusion coefficient approaches 1, the system tends to retain the original model prediction results; when the smoothing fusion coefficient approaches 0, it relies more on the mean of the perturbation sample predictions, thereby mitigating the risk of prediction mutations.

[0103] The perturbation prediction output sequence is averaged to obtain the average prediction result.

[0104] Using the arsenic removal data input sample as input, the current execution decision tree model outputs the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis results, which are recorded as the original prediction results.

[0105] Based on the original prediction results, the average value of the prediction results, and the smoothing fusion coefficient, the original prediction results are smoothed to obtain smoothed prediction results, and a smoothing completion signal is generated simultaneously.

[0106] In a specific embodiment, the smoothed prediction result is specifically represented as follows: , where y is the smoothed prediction result, y1 is the original prediction result, y2 is the average of the prediction results, and d is the smoothing fusion coefficient.

[0107] If the predicted range is greater than or equal to the preset predicted range threshold, the disturbance uncertainty type is recorded as model structure uncertainty, and the fluctuation judgment and processing branch path is recorded as executing lightweight adjustment of model structure.

[0108] If the predicted range is greater than or equal to the predicted range threshold, it indicates a significant jump in the upper and lower bounds of the predicted result. This fluctuation is not caused by input noise, but rather stems from the decision tree model's own structure (such as variable threshold splitting rules) being overly sensitive to the variable currently hitting the sensitive interval. Even a small change in the variable input triggers a model path switch, leading to drastic fluctuations in the output result; that is, the type of disturbance uncertainty is model structural uncertainty. In this case, output smoothing alone cannot fundamentally solve the fluctuation problem. It is necessary to initiate a lightweight adjustment process for the model structure, optimizing the model's response logic to the sensitive interval variable, thereby reducing the risk of prediction jumps at the source and ensuring the reliability of subsequent analysis results.

[0109] Furthermore, lightweight adjustments to the model structure are performed, with the specific execution process as follows:

[0110] A temporary copy model is generated based on the current execution decision tree model.

[0111] In this embodiment, when the system detects that an input sample falls into a sensitive region of the model structure and the fluctuation score of the perturbation prediction result exceeds a threshold, it indicates that the current model has a prediction instability problem in that local region. At this time, directly adjusting the main model may cause the overall prediction ability to degrade. Therefore, it is necessary to keep the main model unchanged to ensure stable prediction of most samples. Based on the current model structure, the model state is copied to generate a temporary copy model for temporary adjustment and experimentation, avoiding the high latency and resource consumption problems caused by full retraining.

[0112] The core purpose of generating a temporary copy model is to perform structural adjustments and adaptive optimizations in locally input-sensitive regions without affecting the stability of the main model, thereby improving the stability of prediction results in model boundary regions or transition regions. A temporary copy model refers to a system that replicates the structure and trained parameter states of the current main model, generating an entity independent of the main model. This temporary copy model maintains structural consistency with the main model, but its hyperparameters controlling complexity and fitting strength can be locally adjusted.

[0113] Use the predicted fluctuation deviation score as the query key to query the model structure fine-tuning parameter set.

[0114] The model structure fine-tuning parameter set includes the minimum number of child node samples increase, the pruning threshold increase, and the split step size reduction.

[0115] In this embodiment, based on the prediction fluctuation deviation score, a pre-built model structure fine-tuning parameter mapping library is invoked. This mapping library is generated based on experimental data from multiple sets of different fluctuation scenarios and internally stores the mapping relationship between different prediction fluctuation deviation score intervals and corresponding structure fine-tuning parameters. The mapping rules follow the principle of positive correlation between prediction fluctuation deviation score and structure fine-tuning intensity: that is, the higher the score (indicating more severe prediction fluctuation), the greater the increase in the minimum number of child nodes (increasing the node splitting threshold and enhancing model stability), the greater the increase in the pruning threshold (strengthening the removal of redundant structures), and the greater the reduction in the splitting step size (slowing down tree growth and avoiding overfitting); conversely, the lower the score, the smaller the corresponding parameter adjustment value.

[0116] Finally, the calculated predicted fluctuation deviation score is used as the query key value. Interval matching is performed in the above mapping library. After locating the interval to which it belongs, the corresponding minimum child node sample number increase value, pruning threshold increase value, and split step size shortening value are extracted, thus completing the query of the model structure fine-tuning parameter set.

[0117] Minimum number of child nodes refers to the minimum number of samples that a child node should contain after each split. If a split results in a child node having fewer than this value, the split will be canceled. This parameter controls whether the model allows split nodes with extremely small sample sizes and is an important means of preventing overfitting.

[0118] The pruning threshold represents the minimum loss reduction required for a node to split. A split will only be performed if it results in a loss reduction greater than this threshold. This parameter essentially controls the "reward threshold" for splits, thus preventing meaningless splits when the benefit is minimal.

[0119] The split step size, or maximum step size, limits the maximum "movement" of each tree during the splitting process, and is particularly useful for offset problems in classification or regression. It helps to slow down the model weight update speed and improve the smoothness of the split.

[0120] Increasing the minimum number of child node samples requires that each split child node must have more samples, which directly suppresses splitting behavior that optimizes only for a very small number of sample points. In boundary regions, since the sample distribution is already sparse, this constraint can prevent the model from making highly sensitive fits in these regions, thereby reducing the possibility of prediction jumps.

[0121] Increasing the pruning threshold forces the model to evaluate a larger benefit before splitting, thus avoiding small, unnecessary splits in sensitive regions of the model boundary. This strategy improves the model's stability within its neighborhood, making it less sensitive to perturbation inputs.

[0122] By limiting the split step size, the magnitude of node weights or split positions can be constrained during each model update, thereby making the generation process of the entire tree structure more stable and preventing certain drastic split rules from appearing near the variable threshold, thus reducing output mutations caused by input perturbations.

[0123] The temporary replica model structure is lightly adjusted based on the model structure fine-tuning parameter set, and an adjustment end signal is generated after the adjustment is completed. At the same time, the temporary replica model that has completed the adjustment is assigned a model state identifier to be tested.

[0124] After generating the replica model, the system fine-tunes its structural hyperparameters, increasing the minimum number of child node samples, raising the pruning threshold, and limiting the maximum split depth. These parameter adjustments help suppress overfitting of the model to perturbation inputs in boundary regions and reduce drastic fluctuations in predictions. Because the replica model is a lightweight adjustment based on the already trained model structure, it eliminates the need to retrain the entire model, thus offering advantages such as speed, low cost, and online operation.

[0125] The disturbance handling effect analysis module is used to analyze the disturbance handling effect after the targeted handling of disturbance fluctuations is completed, and outputs the data analysis results of electronic-grade anhydrous hydrogen fluoride arsenic removal after effective disturbance handling.

[0126] like Figure 2The diagram shown is a flowchart of the disturbance processing effect analysis involved in this embodiment, including two parallel signal receiving paths: When the smoothing completion signal is received, the system analyzes the standard deviation of the original prediction residual based on the disturbance prediction output sequence and the original prediction result, and then analyzes the standard deviation of the smoothed residual based on the disturbance prediction output sequence and the smoothed prediction result. By determining whether the standard deviation of the original prediction residual is greater than the standard deviation of the smoothed residual, if so, the disturbance processing effect analysis result is recorded as effective disturbance processing; otherwise, the disturbance processing effect analysis result is recorded as ineffective disturbance processing. When the adjustment end signal is received, the system inputs the disturbance input sample sequence into the model to be tested and outputs a new disturbance prediction output sequence. Based on the new disturbance prediction output sequence, the prediction fluctuation score is recalculated. If the recalculated prediction fluctuation score is less than the prediction fluctuation score threshold, the disturbance processing effect analysis result is recorded as effective disturbance processing; if the recalculated prediction fluctuation score is greater than or equal to the prediction fluctuation score threshold, the disturbance processing effect analysis result is recorded as ineffective disturbance processing.

[0127] Furthermore, the effect of the disturbance processing was analyzed, and the specific execution process is as follows:

[0128] If the system receives a smooth completion signal, it analyzes the standard deviation of the original prediction residuals based on the perturbation prediction output sequence and the original prediction results.

[0129] Based on the perturbation prediction output sequence and the smoothed prediction results, the standard deviation of the smoothed residuals is analyzed.

[0130] If the standard deviation of the original prediction residuals is greater than the standard deviation of the smoothed residuals, the analysis result of the perturbation treatment effect is recorded as effective perturbation treatment; otherwise, the analysis result of the perturbation treatment effect is recorded as ineffective perturbation treatment.

[0131] If the standard deviation of the original prediction residuals is greater than the standard deviation of the smoothed residuals, it indicates that the perturbation predictions are more dispersed (unstable) under the current model, while the prediction results are more concentrated after smoothing. The original predictions of the current model are unstable within the perturbation range; smoothing filters out jumps and aggregates fluctuations; the smoothed output is more reliable than the original result; therefore, the system considers the perturbation processing effective and can retain the smoothed result for downstream decision-making or output.

[0132] Conversely, this indicates that the predicted disturbance values ​​are already relatively concentrated, and the smoothing process has resulted in the fluctuation range not converging or even expanding. The current smoothing method is not suitable for this distribution structure and has not played a role. If the smoothing result is forcibly adopted, it may introduce bias or information loss. Therefore, the system judges it as invalid disturbance processing and generates an early warning message.

[0133] If the system receives the adjustment end signal, it will input the perturbation input sample sequence into the model to be tested and output a new perturbation prediction output sequence.

[0134] The predicted volatility score is recalculated based on the new perturbation prediction output sequence. If the recalculated predicted volatility score is less than the predicted volatility score threshold, the analysis result of the perturbation treatment effect is recorded as effective perturbation treatment.

[0135] After the system completes model structure adjustment (e.g., adjusting the minimum number of child node samples, pruning parameters, etc.), it re-inputs the previously constructed perturbation input sample sequence into the model to be tested (i.e., the adjusted model), obtaining a new perturbation prediction output sequence. Next, the system recalculates the prediction volatility score of this sequence, which reflects the prediction stability of the adjusted model within the input perturbation range. If the recalculated prediction volatility score is less than the prediction volatility score threshold, it indicates that the model adjustment has significantly improved the prediction jump problem, and the unstable predictions caused by the original input perturbation have been suppressed or mitigated. The current model has higher robustness to sensitive input regions. Therefore, the system can determine that this adjustment is an effective perturbation treatment, and the result can be directly adopted and entered into the output or archiving stage without further smoothing or alarms.

[0136] Furthermore, after effective perturbation treatment, the data analysis results of arsenic removal using electronic-grade anhydrous hydrogen fluoride are output. The specific execution process is as follows:

[0137] After effective perturbation handling, the system performs a model switch, replacing the current decision tree model with the model to be tested, and assigning a status identifier of the current decision tree model to the model to be tested after the replacement. The system also records the switching event, adjustment parameters, score improvement magnitude, and operation timestamp simultaneously for subsequent model version management.

[0138] Input the arsenic removal data as a sample, input the currently executing decision tree model, and output the arsenic removal data analysis results of electronic-grade anhydrous hydrogen fluoride.

[0139] If the recalculated predicted volatility score is greater than or equal to the predicted volatility score threshold, the analysis result of the disturbance processing effect will be recorded as invalid disturbance processing, and an early warning will be generated.

[0140] If the recalculated predicted volatility score is greater than or equal to the predicted volatility score threshold, it indicates that the current model adjustment strategy has failed to improve the predicted volatility in the perturbation zone, and may even exacerbate instability. The model still produces significant output differences for small input perturbations, the structural jump risk has not been eliminated, and the adjustment strategy may not be applicable to this type of perturbation pattern, posing a risk of blind tuning or overfitting correction. Therefore, the system determines that the perturbation handling is invalid and triggers an alarm mechanism, prompting maintenance personnel or system administrators to intervene manually or invoke more advanced strategies (such as model replacement, transfer learning, etc.). This step can prevent the system from repeatedly adjusting in the wrong direction, ensuring the controllability of the overall prediction output and engineering safety.

[0141] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0142] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0143] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0144] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0145] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0146] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0147] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0148] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part 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, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. 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.

[0149] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data analysis system for arsenic removal using an anhydrous hydrogen fluoride in electronic grades based on a decision tree model, characterized in that, include: The variable sensitive interval hit determination module is used to collect the set of electronic-grade anhydrous hydrogen fluoride arsenic removal data variables to form arsenic removal data input samples, and to determine the hit of variable sensitive intervals. When the sensitive interval is not hit, the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis results are output through the currently executed decision tree model. The fluctuation judgment and processing branch path determination module is used to count the number of variables hit when a sensitive interval is hit, thereby generating a disturbance input sample sequence, which is then input into the currently executing decision tree model to obtain the disturbance prediction output sequence, analyze the predicted fluctuation score, determine the fluctuation judgment and processing branch path, and perform targeted processing of disturbance fluctuations. The disturbance handling effect analysis module is used to analyze the disturbance handling effect after the targeted handling of disturbance fluctuations is completed, and output the data analysis results of electronic-grade anhydrous hydrogen fluoride arsenic removal after effective disturbance handling. The process for determining whether the variable sensitive interval has been hit is as follows: The core variable data of the electronic-grade anhydrous hydrogen fluoride arsenic removal process are collected at a preset collection cycle to form the arsenic removal data input sample; Extract the sensitivity interval matching table of each core variable built into the system; The input samples of arsenic removal data are judged based on the sensitive interval matching table of each core variable. If no core variable in the input sample of arsenic removal data falls into its corresponding sensitive interval, the result of the variable sensitive interval hit judgment is recorded as a non-hit sensitive interval. If a core variable in the arsenic removal data falls into its corresponding core variable sensitive interval, the result of the variable sensitive interval hit determination is recorded as the hit sensitive interval, and the core variable is simultaneously recorded as the hit variable.

2. The electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model according to claim 1, characterized in that, The statistical number of hit variables is used to generate a perturbed input sample sequence, and the specific process is as follows: The number of hit variables is counted. If the number of hit variables is 1, a perturbation value is constructed based on the hit variable and the hit variable data, and a perturbation input sample sequence of the hit variable is generated. If the number of hit variables is greater than 1, the Latin hypercube sampling method is used. For each hit variable, it is divided into K equal-width sub-intervals based on its corresponding sensitive interval. Then, random selection is made in each sub-interval to form the perturbation sample set of the hit variable. The perturbation sample sets of each hit variable are combined, shuffled and rematched to generate a linked perturbation sample sequence, which is recorded as the perturbation input sample sequence.

3. The electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model according to claim 2, characterized in that, The specific process for obtaining the perturbation prediction output sequence is as follows: The system loads the decision tree model currently used for the electronic-grade anhydrous hydrogen fluoride arsenic removal process data, calls the model inference interface, and inputs the perturbation input sample sequence as a batch of data into the currently executing decision tree model; The current execution decision tree model outputs the prediction results corresponding to each perturbation input, representing the predicted value of the residual arsenic content under that perturbation input, and all output results are combined into a prediction sequence, which is denoted as the perturbation prediction output sequence.

4. The electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model according to claim 1, characterized in that, The specific process for analyzing and predicting the volatility score is as follows: Based on the perturbation prediction output sequence, the maximum and minimum values ​​of the output prediction results are extracted. The difference between the maximum and minimum values ​​of the prediction results is processed to obtain the prediction result range, which is used to measure the upper and lower bound fluctuation range of the prediction results. Calculate the standard deviation of the perturbation prediction output sequence, denoted as the prediction result standard deviation, which is used to measure the overall dispersion of the prediction results; The perturbation prediction output sequence is fitted to a one-dimensional linear trend line, and the root mean square error between the perturbation prediction output sequence and the fitted value is calculated and denoted as the prediction result trend residual score. The prediction result trend residual score is used to capture nonlinear jump or trend change behavior. By setting a scaling factor, the prediction result range, prediction result standard deviation, and prediction result trend residual score are scaled down, and the maximum value after scaling down is taken as the prediction fluctuation score.

5. The electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model according to claim 1, characterized in that, The specific analysis process for determining the fluctuation judgment and processing branch path is as follows: Extract the predicted volatility score threshold; If the predicted fluctuation score is less than the predicted fluctuation score threshold, the fluctuation judgment processing branch path is recorded as directly taking the arsenic removal data input sample as input, and outputting the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis results based on the current execution decision tree model. If the predicted volatility score is greater than or equal to the predicted volatility score threshold, then a disturbance uncertainty type analysis is performed. If the predicted result range is less than the preset predicted result range threshold, then the disturbance uncertainty type is recorded as input noise uncertainty, and the volatility judgment and processing branch path is recorded as output smoothing. If the predicted range is greater than or equal to the preset predicted range threshold, the disturbance uncertainty type is recorded as model structure uncertainty, and the fluctuation judgment and processing branch path is recorded as executing lightweight adjustment of model structure.

6. The electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model according to claim 5, characterized in that, The smooth execution output is analyzed in the following process: The difference between the predicted volatility score and the predicted volatility score threshold is processed to obtain the predicted volatility deviation score; Using the predicted fluctuation deviation score as the query key, query the smoothing fusion coefficient; The average value of the prediction results is obtained by averaging the perturbation prediction output sequence. Using the arsenic removal data input sample as input, the current execution decision tree model outputs the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis results, which are recorded as the original prediction results; Based on the original prediction results, the average value of the prediction results, and the smoothing fusion coefficient, the original prediction results are smoothed to obtain smoothed prediction results, and a smoothing completion signal is generated simultaneously.

7. The electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model according to claim 5, characterized in that, The execution model structure is lightweight and adjusted; the specific execution process is as follows: Based on the current execution decision tree model, generate a temporary copy model; Use the predicted fluctuation deviation score as the query key to query the model structure fine-tuning parameter set; The model structure fine-tuning parameter set includes the minimum number of child node samples increase value, the pruning threshold increase value, and the split step size reduction value; The temporary replica model structure is lightly adjusted based on the model structure fine-tuning parameter set, and an adjustment end signal is generated after the adjustment is completed. At the same time, the temporary replica model that has completed the adjustment is assigned a model state identifier to be tested.

8. The electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model according to claim 1, characterized in that, The specific execution process for analyzing the effect of perturbation processing is as follows: If the system receives a smooth completion signal, it analyzes the standard deviation of the original prediction residuals based on the perturbation prediction output sequence and the original prediction results. Based on the perturbation prediction output sequence and the smoothed prediction results, the standard deviation of the smoothed residuals is analyzed. If the standard deviation of the original prediction residuals is greater than the standard deviation of the smoothed residuals, the analysis result of the perturbation treatment effect is recorded as effective perturbation treatment; otherwise, the analysis result of the perturbation treatment effect is recorded as ineffective perturbation treatment. If the system receives a regulation termination signal, it will input the perturbation input sample sequence into the model to be tested and output a new perturbation prediction output sequence; The predicted volatility score is recalculated based on the new perturbation prediction output sequence. If the recalculated predicted volatility score is less than the predicted volatility score threshold, the analysis result of the perturbation treatment effect is recorded as effective perturbation treatment. If the recalculated predicted volatility score is greater than or equal to the predicted volatility score threshold, the analysis result of the disturbance processing effect will be recorded as invalid disturbance processing, and an early warning will be generated.

9. The electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis system based on a decision tree model according to claim 8, characterized in that, After effective perturbation processing, the electronic-grade anhydrous hydrogen fluoride arsenic removal data analysis results are output. The specific execution process is as follows: After effective perturbation handling, the system performs model switching, replacing the current execution decision tree model with the model to be tested, and assigning the current execution decision tree model status identifier to the model to be tested after the replacement, and synchronously recording the switching event, adjustment parameters, score improvement range and operation timestamp for subsequent model version management; Input the arsenic removal data as a sample, input the currently executing decision tree model, and output the arsenic removal data analysis results of electronic-grade anhydrous hydrogen fluoride.

Images