A method and system for predicting the life of a chemical solution of an alkali etching cleaning machine based on big data

Abstract

The present application relates to a kind of based on big data's alkali throwing cleaning machine liquid life prediction method and system, it is related to the field of alkali throwing cleaning machine liquid life prediction, wherein based on big data's alkali throwing cleaning machine liquid life prediction method includes: the steps of building multidimensional feature parameter acquisition system, constructing liquid life prediction model, model training verification and self-adapting correction, real-time prediction and failure early warning, generating liquid replacement or supplement suggestion, based on big data's alkali throwing cleaning machine liquid life system includes, acquisition module, data processing module, prediction module, training module, early warning module, feedback module and data storage module: by collecting effective alkali concentration, silicon ion concentration, impurity ion concentration, pH, temperature, viscosity, process condition and other multidimensional feature parameters, not only rely on single pH value or alkali concentration, comprehensively reflect the real attenuation state of liquid, improve prediction accuracy.

Description

Technical Field

[0001] This invention relates to the technical field of predicting the lifespan of chemicals used in alkaline polishing machines, and in particular to a method and system for predicting the lifespan of chemicals used in alkaline polishing machines based on big data. Background Technology

[0002] In industries such as semiconductor manufacturing and photovoltaic silicon wafer processing, alkaline polishing is one of the core processes. Its main function is to polish, deoxidize, and remove impurities from the workpiece surface using alkaline solutions (such as sodium hydroxide and potassium hydroxide solutions), ensuring the smoothness and cleanliness of the workpiece surface and laying the foundation for subsequent processes. The performance of the alkaline polishing solution directly determines the cleaning effect and product yield. As the cleaning operation progresses, the effective alkaline component in the solution is continuously consumed, and silicon, metallic impurities, etc., on the workpiece surface dissolve into the solution, leading to a decrease in solution concentration and an increase in impurity content. When the performance of the solution deteriorates to a certain level, it will be unable to meet the cleaning requirements, resulting in problems such as workpiece surface scratches, uneven polishing, and excessive residual impurities. At this point, the solution needs to be replaced promptly.

[0003] Currently, the industry relies primarily on manual experience to determine the lifespan of alkaline polishing cleaning solutions. Operators judge whether the solution needs replacement based on the batch quantity of workpieces being cleaned, cleaning time, changes in the solution's appearance (such as color and turbidity), and past experience. This method has significant drawbacks: First, it is highly subjective, with significant differences in experience among operators, easily leading to judgment errors. This can result in either premature solution replacement leading to waste or delayed replacement causing a decrease in product yield. Second, it cannot achieve real-time monitoring. Manual inspection intervals are long, making it difficult to capture dynamic changes in solution performance, especially in large-scale, continuous production where solution performance may rapidly decline in a short period, which cannot be detected in time by manual judgment. Third, it lacks quantitative indicators, making it impossible to accurately predict the remaining lifespan of the solution and develop a reasonable solution replacement plan, thus affecting production efficiency. Summary of the Invention

[0004] In order to achieve accurate prediction and real-time monitoring of the lifespan of alkaline polishing cleaning solutions, this invention provides a method and system for predicting the lifespan of alkaline polishing cleaning solutions based on big data.

[0005] In a first aspect, the present invention provides a method for predicting the lifespan of alkaline polishing cleaning solutions based on big data, employing the following technical solution: A method for predicting the lifespan of alkaline polishing cleaning solutions based on big data includes: S1. Establish a multi-dimensional feature parameter acquisition system: Establish a multi-dimensional feature parameter acquisition system for the alkaline polishing machine solution to collect dynamic feature parameters of the solution at different working periods. The dynamic feature parameters include chemical feature parameters, physical feature parameters and process condition parameters, comprehensively capturing various factors affecting the life of the solution and providing sufficient data support for subsequent prediction models. S2: Feature parameter preprocessing: The collected dynamic feature parameters are preprocessed, including outlier removal, data standardization, feature selection, removal of redundant and outlier data, elimination of the influence of units, and obtaining an effective set of feature parameters for model training, thereby improving the training efficiency and prediction accuracy of the model. S3: Constructing a drug liquid life prediction model: The prediction model is a coupled model of an improved BP neural network and a drug liquid kinetic decay model. The improved BP neural network is used to initially predict the remaining life of the drug liquid, and the kinetic decay model is used to correct the initial prediction results. Combining the advantages of the two models improves the prediction accuracy and solves the problem of large prediction deviation of a single model. S4: Model Training, Validation and Adaptive Correction: Based on historical data and real-time collected effective feature parameters, the prediction model is trained, validated and adaptively corrected to ensure that the model prediction accuracy meets the preset requirements, while improving the model's adaptability and stability, and enabling long-term stable operation. S5: Real-time prediction and failure warning: Real-time acquisition of dynamic characteristic parameters of the alkali blasting cleaning liquid, preprocessing according to the method in step S2 to obtain effective characteristic parameters, inputting them into the trained prediction model, the model outputs the remaining life prediction value Lfinal of the liquid and failure warning signal, realizing real-time monitoring of the liquid life and failure warning. S6: Generate suggestions for replacing or replenishing the solution: Based on the predicted remaining lifespan of the solution and real-time data on effective alkali concentration and silicon ion concentration, generate targeted suggestions for replacing or replenishing the solution to guide operators in subsequent operations and achieve intelligent management of the solution.

[0006] Optionally, in step S1, the chemical characteristic parameters include: effective alkali concentration, silicon ion concentration, impurity ion concentration, surfactant concentration, and pH value; wherein the effective alkali concentration is the mass concentration of sodium hydroxide or potassium hydroxide, the silicon ion concentration is the concentration of silicate ions generated by the reaction of the solution with the silicon wafer, and the impurity ion concentration includes the concentrations of iron ions, copper ions, and sodium ions; the physical characteristic parameters include the temperature, viscosity, density, and surface tension of the solution; and the process condition parameters include the material of the workpiece being cleaned, the number of workpiece batches, the single cleaning time, the set value of the cleaning temperature, the circulation speed of the solution, and the frequency and amount of replenishment of the solution.

[0007] Optionally, the specific process of step S2 is as follows: S21. Outlier Removal: The Grubbs criterion is used to detect outliers in the collected dynamic feature parameters and remove outlier data that exceed 3 times the standard deviation to ensure data reliability. S22. Data Standardization: The min-max standardization method is used to map the feature parameters after removing outliers to the [0, 1] interval, eliminating the influence of different dimensions of parameters. The standardization formula is: ; Here, x represents the standardized feature parameter values, and x represents the original feature parameter values. This is the minimum value of the characteristic parameter. This is the maximum value of the feature parameter; S23. Feature Selection: The mutual information method is used to calculate the mutual information value between each feature parameter and the lifespan of the liquid. Feature parameters with mutual information values ​​greater than the preset threshold (0, 3) are selected to form an effective feature parameter set. Redundant parameters with weak correlation to the lifespan of the liquid are eliminated to reduce the computational complexity of the model.

[0008] Optionally, in step S3, the improved BP neural network structure includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is the same as the number of parameters in the effective feature parameter set. The hidden layer has 2-3 layers, with each layer containing 1.5-2 times the number of neurons in the input layer. The output layer has 1 neuron, and the output is the predicted remaining lifespan of the drug solution. The improved BP neural network uses momentum gradient descent to optimize weight updates, introduces a dropout mechanism to prevent overfitting, and employs an adaptive learning rate adjustment strategy, dynamically adjusting the learning rate based on the training error. The learning rate adjustment formula is: ; in, Let be the learning rate for the (k+1)th iteration. Let be the learning rate for the k-th iteration. Let be the training error of the k-th iteration. This represents the training error of the (k+1)th iteration; The drug solution kinetic decay model is constructed based on the consumption law of the active ingredient in the drug solution, and its expression is:

[0009] Where L(t) is the remaining lifespan of the drug solution at time t. Let λ be the initial lifetime of the drug solution, λ be the decay coefficient of the effective components of the drug solution, t be the working time of the drug solution, α be the correction coefficient (range 0.05-0.15), and n be the number of impurity ions. Let be the concentration of the i-th impurity ion at time t. The weight of the influence of the i-th impurity ion on the lifespan of the drug solution; The output formula of the coupled model is: ; in, This is the final predicted value for the remaining lifespan of the drug solution. The values ​​are the initial predictions of the improved BP neural network, where ω is the weight coefficient (ranging from 0.6 to 0.8) and L(t) is the correction value of the dynamic decay model.

[0010] Optionally, the specific process of step S4 is as follows: S41. Data partitioning: Divide the effective feature parameter set into training set, validation set and test set in a ratio of 7:2:1. The training set is used for model training, the validation set is used to adjust model parameters and the test set is used to verify the model prediction accuracy. S42. Model Training: Input the training set into the improved BP neural network, initialize the model parameters (weights, threshold, learning rate, momentum factor), set the number of training iterations (1000-5000 times) and the training error threshold (≤0.001), and use the momentum gradient descent method to update the weights until the training error reaches the preset threshold or the maximum number of training iterations is reached, thus completing the initial training of the BP neural network. S43. Model calibration: Substitute the initial predicted values ​​output by the trained BP neural network into the drug kinetic decay model for calibration to obtain the prediction results of the coupled model. Use the validation set to adjust the weight coefficient ω, correction coefficient α and decay coefficient λ of the coupled model to minimize the prediction error of the validation set. S44. Model Validation: Input the test set into the calibrated coupled model and calculate the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) between the predicted and actual drug lifespan. 2 If R 2 If the accuracy is ≥0.95, MAE≤0.5h, and RMSE≤0.8h, then the model meets the preset accuracy requirements; if not, return to step S42, adjust the model parameters, and retrain. S45. Adaptive correction: Real-time acquisition of dynamic characteristic parameters and actual lifespan data of the liquid medicine, periodic (every 24 hours) fine-tuning of the coupled model, updating model parameters, adapting to changes in liquid medicine composition and operating condition fluctuations, and ensuring the long-term prediction accuracy of the model.

[0011] Optionally, in step S5, the failure warning signal is divided into three levels: Level 1 warning (remaining liquid life ≥ 2h), prompting the operator to pay attention to the liquid status and strengthen parameter monitoring; Level 2 warning (1h ≤ remaining liquid life < 2h), prompting the operator to prepare to replace or replenish the liquid; Level 3 warning (remaining liquid life < 1h), immediately issuing an alarm signal, prompting the operator to stop the cleaning operation and replace the liquid; the warning signal is synchronously output through the audible and visual alarm device and the host computer system, and the warning time and corresponding characteristic parameters are recorded at the same time; In step S6, the suggestion to replace or replenish the solution is generated based on the predicted remaining lifespan of the solution and real-time data of the effective alkali concentration and silicon ion concentration: when the remaining lifespan of the solution is <1h or the effective alkali concentration is lower than the preset threshold (1.5%) or the silicon ion concentration is higher than the preset threshold (5000ppm), a suggestion to replace the solution immediately is generated; when 1h ≤ remaining lifespan of the solution <2h and the effective alkali concentration is within the preset threshold range (1.5%-3%), a suggestion to replenish the solution is generated. The replenishment amount is calculated based on the difference between the effective alkali concentration and the standard concentration and the total amount of solution. The formula for calculating the replenishment amount is as follows: ; in, This is to replenish the amount of medicine. This refers to the total volume of the cleaning solution in the alkaline polishing machine. The effective alkali standard concentration of the drug solution is 3%–5%. This represents the real-time effective alkali concentration of the drug solution. To replenish the effective alkali concentration of the solution (10%-15%), the generated suggestions also include the time points for solution replacement or replenishment, operating procedures and precautions, which are simultaneously pushed to the operator's terminal.

[0012] Optionally, in step S1, the characteristic parameters are collected every 5-10 minutes. The effective alkali concentration is collected by titration combined with an online concentration detector, the silicon ion concentration is collected by spectrophotometry, the impurity ion concentration is collected by inductively coupled plasma optical emission spectrometry (ICP-OES), the pH value is collected by an online pH meter, and the temperature, viscosity, density, and surface tension are collected by online temperature sensors, viscosity sensors, density sensors, and surface tension sensors, respectively. The process parameters are read in real time by the PLC control system of the alkali polishing cleaning machine.

[0013] Optionally, step S7, data storage and traceability, is also included: storing the collected dynamic characteristic parameters, the input and output data of the prediction model, the drug solution replacement records, the early warning records, and other data into the database for a period of not less than one year, so as to facilitate subsequent traceability analysis, model optimization, and process improvement.

[0014] Secondly, the present invention provides a big data-based system for predicting the lifespan of chemicals used in alkaline polishing machines, characterized in that it includes: The data acquisition module is electrically connected to the input of the data processing module. It is used to build a multi-dimensional characteristic parameter acquisition system for the alkaline polishing cleaning machine solution. The system collects the dynamic characteristic parameters of the solution at different working times. The dynamic characteristic parameters include chemical characteristic parameters, physical characteristic parameters, and process condition parameters. Data processing module: The output end is electrically connected to the input end of the prediction module. It is used to preprocess the dynamic feature parameters collected by the acquisition module, including outlier removal, data standardization, feature filtering, removal of redundant and outlier data, elimination of the influence of units, and obtaining an effective set of feature parameters for model training. Prediction module: The output terminal is electrically connected to the input terminal of the training module and is used to build a prediction model. The prediction model is a coupled model of an improved BP neural network and a drug liquid kinetic decay model. The improved BP neural network is used to initially predict the remaining life of the drug liquid, and the kinetic decay model is used to correct the initial prediction results. Training module: The output end is electrically connected to the input end of the early warning module. Based on historical data and real-time collected effective feature parameters, it trains, validates, and adaptively corrects the prediction model. Early warning module: The output end is electrically connected to the input end of the feedback module. It is used to collect the dynamic characteristic parameters of the alkali blasting cleaning liquid in real time, preprocess it according to the method of the data processing module to obtain effective characteristic parameters, and input them into the trained prediction model. The model outputs the remaining life of the liquid and failure early warning signal. Feedback module: The output end is electrically connected to the input end of the data storage module. Based on the predicted remaining lifespan of the drug solution and real-time data of effective alkali concentration and silicon ion concentration, it generates targeted suggestions for drug solution replacement or replenishment. Data storage module: Stores dynamic feature parameters collected by the acquisition module, input and output data of the prediction model, drug solution replacement records, early warning records, and other data in the database.

[0015] In summary, the present invention has at least one of the following beneficial technical effects: 1. This invention comprehensively reflects the true decay state of the drug solution by collecting multi-dimensional characteristic parameters such as effective alkali concentration, silicon ion concentration, impurity ion concentration, pH, temperature, viscosity, and process conditions, rather than relying solely on a single pH value or alkali concentration. By combining an improved BP neural network with a kinetic decay model for prediction, it avoids the problems of insufficient fitting by traditional linear models and poor generalization of single neural networks.

[0016] 2. This invention features a three-tiered early warning mechanism, capable of issuing warnings 1-2 hours before the chemical solution becomes ineffective. This prevents issues such as uneven polishing of silicon wafers / crystals, dirt residue, and surface defects caused by the degradation of the chemical solution's performance. Compared to manual, timed sampling and experience-based judgment, this invention can capture sudden changes in the chemical solution's state in real time, effectively reducing the generation of defective products and lowering rework and scrap costs. It is particularly suitable for continuous, high-volume production lines.

[0017] 3. Traditional production often results in significant waste due to premature replacement of the cleaning solution, or quality risks due to delayed replacement. This invention can accurately determine the remaining lifespan and optimal replacement time, and automatically calculate the replenishment amount based on the real-time concentration. While ensuring cleaning quality, it extends the effective usage time of the cleaning solution by 10%–20%, reduces the amount of alkali solution prepared and waste liquid discharged, and significantly reduces raw material costs and hazardous waste treatment costs.

[0018] 4. This invention adaptively fine-tunes the coupled model by periodically introducing newly collected data, adapting to the effects of changes in different seasons, workpiece materials, and process parameters, thus solving the problem of "one-time training, long-term deviation" in traditional models. The model maintains high accuracy even under fluctuating operating conditions, eliminating the need for frequent manual recalibration, and is highly versatile and compatible with different models of alkaline polishing cleaning machines.

[0019] 5. The system automatically generates suggestions for chemical solution replacement / replenishment, including replenishment volume, timing, and steps, reducing reliance on operator experience. Simultaneously, a complete database is established for long-term storage of parameters, warnings, and replacement records, supporting one-click traceability by time, batch, and equipment number. This meets the quality traceability and system audit requirements of the photovoltaic and semiconductor industries, facilitating process optimization and anomaly tracing. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of Embodiment 1 of this application; Figure 2 This is a schematic diagram of the system structure of Embodiment 2 of this application. Detailed Implementation

[0021] The following combination Figures 1 to 2 The present invention will be described in further detail below.

[0022] Example 1: This application provides a method for predicting the lifespan of alkaline polishing cleaning solutions based on big data, referring to... Figure 1 A method for predicting the lifespan of alkaline polishing cleaning solutions based on big data includes the following steps; S1. Establish a multi-dimensional feature parameter acquisition system: Establish a multi-dimensional feature parameter acquisition system for the alkaline polishing machine solution to collect dynamic feature parameters of the solution at different working periods. The dynamic feature parameters include chemical feature parameters, physical feature parameters, and process condition parameters, comprehensively capturing various factors affecting the lifespan of the solution and providing sufficient data support for subsequent prediction models.

[0023] Among them, chemical characteristic parameters are the core parameters that determine the performance of the drug solution, directly reflecting the content of the effective components and the degree of contamination, including: Effective alkali concentration: The mass concentration of sodium hydroxide (NaOH) or potassium hydroxide (KOH) in the cleaning solution is the core component that enables the cleaning solution to achieve polishing and deoxidation functions. As the cleaning operation proceeds, the effective alkali will react with the silicon and oxide layer on the surface of the workpiece and be continuously consumed, and the concentration will gradually decrease. When the concentration is lower than the preset threshold, the cleaning solution cannot meet the cleaning requirements. Silicon ion concentration: Silicate ions (SiO3) generated by the reaction of the chemical solution with the silicon wafer (or silicon-containing workpiece). 2- The higher the concentration of silicon ions, the more reaction products accumulate in the solution, which will significantly reduce the polishing effect of the solution and increase the viscosity of the solution, affecting the circulation and cleaning uniformity of the solution. Impurity ion concentration: including iron ions (Fe) 3+ ), copper ions (Cu) 2+ Sodium ions (Na) + These impurity ions mainly come from residues on the workpiece surface, equipment wear and tear, and environmental pollution. Impurity ions can react with the active ingredients in the solution, accelerating the solution's failure, and can also cause impurity residues on the workpiece surface, affecting product quality. Surfactant concentration: The concentration of surfactants (such as nonylphenol anionic surfactants) added to the solution. Surfactants can reduce the surface tension of the solution, allowing the solution to better and more evenly contact and react with the workpiece surface. They also have an antifoaming effect, reducing the impact of bubbles generated during the reaction on the cleaning effect. A decrease in their concentration will lead to a decrease in cleaning uniformity. pH value: The acidity or alkalinity of the solution directly affects its reactivity. The pH value of alkaline scouring solutions is usually between 12 and 14. As the effective alkali is consumed and reaction products accumulate, the pH value will gradually decrease. When the pH value is below 12, the reactivity of the solution will be significantly reduced.

[0024] Physical characteristic parameters include: Temperature: The working temperature of the cleaning solution. The optimal temperature for alkaline blasting is usually 55-65℃. Too high a temperature will accelerate the decomposition and volatilization of the active ingredients in the cleaning solution, while too low a temperature will reduce the reaction rate and affect the cleaning efficiency. Temperature fluctuations will directly affect the lifespan of the cleaning solution. Viscosity: The viscosity of the solution. Viscosity is mainly affected by the concentration of silicon ions and surfactants. Increased viscosity will lead to a decrease in the circulation speed of the solution, a decrease in the uniformity of cleaning, and an increase in the consumption of the solution. Density: The ratio of the mass to the volume of the medicinal solution. As the effective alkali is consumed and impurities accumulate, the density of the medicinal solution will gradually change. Abnormal fluctuations in density reflect changes in the composition of the medicinal solution. Surface tension: the intermolecular force on the surface of a cleaning solution. The lower the surface tension, the easier it is for the cleaning solution to wet the workpiece surface, resulting in a better cleaning effect. Changes in surface tension are closely related to the concentration of surfactants and the content of impurities.

[0025] Process operating parameters include: The materials of the workpieces to be cleaned: such as photovoltaic silicon wafers, semiconductor wafers, metal workpieces, etc. Different materials of workpieces react with the cleaning solution at different rates, and the consumption rate of the effective components of the cleaning solution is also different. Batch quantity of workpieces: The number of workpieces cleaned per unit time. The more batches there are, the faster the chemical solution is consumed and the faster impurities accumulate. Single cleaning time: The cleaning time for each batch of workpieces. The longer the cleaning time, the more fully the reaction between the solution and the workpiece, and the more effective ingredients are consumed. Cleaning temperature setting: The preset working temperature of the alkaline polishing machine. The higher the set temperature, the stronger the reaction activity of the chemical solution and the faster the active ingredients are consumed. Chemical circulation speed: The speed at which the chemical solution circulates within the cleaning machine. The faster the circulation speed, the more uniform the contact between the chemical solution and the workpiece, and the more uniform the consumption of the effective ingredients. However, it will also accelerate the evaporation and decomposition of the chemical solution. Frequency and dosage of replenishing the cleaning solution: During the cleaning process, fresh cleaning solution may be added periodically to maintain its performance. The frequency and dosage of replenishment will directly affect the lifespan and performance stability of the cleaning solution.

[0026] The characteristic parameters are collected every 5-10 minutes to ensure real-time capture of the dynamic changes in the drug solution. Specifically: the effective alkali concentration is collected using titration combined with an online concentration detector; the titration method is used to periodically calibrate the accuracy of the online detector to ensure data accuracy. The silicon ion concentration is collected using spectrophotometry, calculating the silicon ion concentration by detecting the absorbance of the drug solution at a specific wavelength. The impurity ion concentration is collected using inductively coupled plasma optical emission spectrometry (ICP-OES), a method with high detection accuracy that can simultaneously detect the concentrations of multiple impurity ions. The pH value is collected using an online pH meter to monitor the acidity and alkalinity changes of the drug solution in real time. Temperature, viscosity, density, and surface tension are collected using online temperature sensors, viscosity sensors, density sensors, and surface tension sensors, respectively. The measurement accuracy of the sensors meets the preset requirements (temperature ±0.5℃, viscosity ±0.01mPa·s, density ±0.001g / cm³).3 (Surface tension ±0.1mN / m); process parameters are read in real time by the PLC control system of the alkaline polishing machine to ensure the real-time performance and accuracy of the data.

[0027] S2. Feature parameter preprocessing: This includes outlier removal, data standardization, and feature selection. It removes redundant and outlier data, eliminates the influence of units, and obtains an effective set of feature parameters for model training, thereby improving the training efficiency and prediction accuracy of the model.

[0028] S21. Outlier Removal: Due to sensor malfunctions, environmental interference, operational errors, etc., the collected feature parameters may contain outliers. These outliers can severely affect the training effect of the model. Therefore, the Grubbs criterion is needed to detect and remove outliers. The specific steps of the Grubbs criterion are as follows: Calculate the average value of a certain characteristic parameter. and standard deviation s; Calculate the absolute value of the deviation of each data point from the mean. ; Calculate the Grubbs statistic G: ; Based on the sample size n and the confidence level (95% confidence level is used in this invention), find the Grubbs critical value Gα; If G > Gα, then the data point is an outlier and should be removed; otherwise, the data point should be retained.

[0029] The Grubbs criterion can effectively eliminate outlier data exceeding three standard deviations, ensuring the reliability and authenticity of the data.

[0030] S22. Data Standardization: Since feature parameters of different dimensions have different units (e.g., concentration is expressed as %, temperature as ℃, and viscosity as mPa·s), directly inputting them into the model can cause the model to become overly sensitive to parameters with large units, affecting the model's training performance and prediction accuracy. Therefore, the min-max standardization method is needed to map the feature parameters after removing outliers to the [0,1] interval, eliminating the influence of units. The standardization formula is: ;in, Here, x represents the standardized feature parameter values, and x represents the original feature parameter values. This is the minimum value of the characteristic parameter. This is the maximum value of the feature parameter.

[0031] For example, the original range of effective alkali concentration is 1.5%–5%. If the effective alkali concentration at a certain moment is 3%, the minimum value of this parameter is 1.5%, and the maximum value is 5%. Then the standardized value is (3-1.5) / (5-1.5)≈0.4286. Through standardization, all feature parameters are unified to the same numerical range, which facilitates model training and calculation.

[0032] S23. Feature Filtering: Among the collected dynamic feature parameters, some parameters have a weak correlation with the lifespan of the drug solution. Inputting all parameters into the model would increase computational complexity, leading to overfitting and reduced prediction accuracy. Therefore, a mutual information method is needed to calculate the mutual information value between each feature parameter and the lifespan of the drug solution, filtering out highly correlated feature parameters to form an effective feature parameter set. The mutual information method measures the degree of correlation between two random variables; a larger mutual information value indicates a stronger correlation between the feature parameter and the lifespan of the drug solution, and vice versa. In this invention, a preset mutual information value threshold of 0.3 is used to filter out feature parameters with a mutual information value greater than 0.3, while redundant parameters with a mutual information value less than or equal to 0.3 are removed. For example, through mutual information calculation, it was found that the mutual information values ​​of effective alkali concentration, silicon ion concentration, pH value, temperature, batch quantity of workpieces and liquid life are 0.85, 0.82, 0.78, 0.65 and 0.52 respectively, all greater than 0.3. Therefore, these parameters are included in the effective feature parameter set. On the other hand, the mutual information values ​​of sodium ion concentration and surface tension and liquid life are 0.28 and 0.25 respectively, less than 0.3. Therefore, they are removed, thereby reducing the computational complexity of the model and improving the training efficiency of the model.

[0033] S3. Constructing a drug liquid life prediction model: Constructing a drug liquid life prediction model, which is a coupled model of an improved BP neural network and a drug liquid kinetic decay model. The improved BP neural network is used to initially predict the remaining life of the drug liquid, and the kinetic decay model is used to correct the initial prediction results. Combining the advantages of the two models improves the prediction accuracy and solves the problem of large prediction deviation of a single model.

[0034] S31: Construction of the Improved BP Neural Network: The BP neural network (Back Propagation Neural Network) is a multi-layer feedforward neural network with strong nonlinear fitting capabilities. It can handle complex multi-factor mapping relationships and is suitable for nonlinear prediction problems such as drug lifespan, which are affected by multiple factors. However, traditional BP neural networks have drawbacks such as slow convergence speed, susceptibility to overfitting, and poor training results due to a fixed learning rate. Therefore, this invention improves the traditional BP neural network. The specific improvement measures are as follows: S311: Network Structure Optimization: The improved BP neural network structure includes an input layer, hidden layers, and an output layer. The number of neurons in the input layer is consistent with the number of parameters in the effective feature parameter set. For example, if the effective feature parameter set includes five parameters: effective alkali concentration, silicon ion concentration, pH value, temperature, and batch quantity of workpieces, then the number of neurons in the input layer is 5. The hidden layer has 2-3 layers, with the number of neurons in each layer being 1.5-2 times the number of neurons in the input layer. For example, when the number of neurons in the input layer is 5, the first hidden layer has 8 neurons (5 × 1.6), and the second hidden layer has 10 neurons (8 × 1.25). The hidden layer uses the ReLU activation function to solve the gradient vanishing problem of the traditional Sigmoid activation function. The ReLU activation function expression is: The output layer has 1 neuron and outputs the predicted remaining lifespan of the drug solution. The output layer uses a linear activation function to ensure the continuity of the predicted value.

[0035] S312: Optimized Weight Update Method: The momentum gradient descent method is used to optimize the weight update. Based on the traditional gradient descent method, the momentum gradient descent method introduces a momentum factor to reduce oscillations during training and accelerate convergence. The weight update formula is as follows: ;in, The weight for the (k+1)th iteration. Let the weight be the weight for the k-th iteration. Let η be the weights for the (k-1)th iteration, and η be the learning rate. Let α be the weight gradient of the k-th iteration, and α be the momentum factor (ranging from 0.8 to 0.9).

[0036] S313: Introducing dropout mechanism: To prevent model overfitting, a dropout mechanism is introduced in the hidden layer. During each training session, some hidden layer neurons are randomly dropped (the dropout ratio is 0.2-0.3), so that the model does not overly rely on a certain neuron and improves the model's generalization ability.

[0037] S314: Adaptive Learning Rate Adjustment: An adaptive learning rate adjustment strategy is adopted, dynamically adjusting the learning rate based on the training error. When the training error decreases, the learning rate is appropriately increased to accelerate convergence; when the training error increases, the learning rate is appropriately decreased to avoid model oscillation. The learning rate adjustment formula is: ;in, Let be the learning rate for the (k+1)th iteration. The learning rate for the k-th iteration (the initial learning rate ranges from 0.001 to 0.01). Let be the training error of the k-th iteration. Let be the training error of the (k+1)th iteration. Through the above improvements, the convergence speed of the improved BP neural network is increased by more than 40% compared with the traditional BP neural network, and its generalization ability is significantly enhanced. It can effectively handle complex nonlinear problems in drug liquid life prediction and achieve preliminary prediction of the remaining life of the drug liquid.

[0038] S32: Construction of the Drug Liquid Kinetic Decay Model: The decay process of drug liquid lifetime is essentially a kinetic process of active ingredient consumption and impurity accumulation, following certain decay laws. Therefore, this invention constructs a drug liquid kinetic decay model based on the consumption law of active ingredients in the drug liquid to correct the preliminary prediction results of the improved BP neural network and improve prediction accuracy. The core idea of ​​the drug liquid kinetic decay model is that the remaining lifetime of the drug liquid is closely related to the decay rate of active ingredients and the accumulation rate of impurity ions. The faster the active ingredients decay and the more impurity ions accumulate, the shorter the remaining lifetime of the drug liquid. Combining the principles of chemical reaction kinetics, the expression of the constructed drug liquid kinetic decay model is: Where L(t) is the remaining lifespan of the liquid medicine at time t (unit: h), that is, the time during which the liquid medicine can still work normally from time t. λ represents the initial lifetime of the cleaning solution (in hours), preset according to the solution formulation, initial concentration, and process requirements, with a range of 8-24 hours. For example, the initial lifetime of the alkaline polishing cleaning solution for photovoltaic silicon wafers is typically 12-16 hours, while that for semiconductor wafers is typically 16-24 hours. λ is the attenuation coefficient of the effective components in the cleaning solution (in hours). -1 This reflects the consumption rate of the effective alkaline components, and is obtained by fitting historical data, with a value range of 0.02-0.08h. -1 The faster the effective alkali concentration is consumed, the larger the λ value. For example, when cleaning a large number of workpieces in batches at high temperatures, the λ value is typically 0.06-0.08h. -1 When the number of workpieces being cleaned in small batches and the temperature is low, the λ value should be between 0.02 and 0.04 h. -1 t represents the working time of the medicine solution (unit: h), which is the cumulative working time from the time the medicine solution is replaced until the current time; α is a correction coefficient, used to correct the degree of influence of impurity ion accumulation on the lifespan of the medicine solution, with a value range of 0.05-0.15, calibrated based on historical data. The greater the influence of impurity ions on the lifespan of the medicine solution, the larger the value of α; n represents the number of impurity ion types. In this invention, the three main impurity ions are iron ions, copper ions, and sodium ions, so n=3. The concentration of the i-th impurity ion at time t (unit: ppm) is collected in real time by an ICP-OES detector. The weight of the i-th impurity ion on the lifespan of the cleaning solution is set according to the degree of influence of each impurity ion on the cleaning effect of the cleaning solution, where iron ions are included. (Iron ions react with effective alkali to form precipitates, accelerating the deterioration of the drug solution and having the greatest impact), copper ions (Copper ions can cause copper contamination on the workpiece surface, affecting product quality), sodium ions (Sodium ions have a relatively small impact on the properties of the drug solution), other impurity ions The sum of the influence weights of all impurity ions is 1, where: 0.4 + 0.3 + 0.2 + 0.1 = 1. The sum of the weights is strictly 1, which is consistent with the model setting.

[0039] This kinetic decay model can accurately fit the decay law of the effective components of the drug solution and the impact of impurity ion accumulation on the lifespan of the drug solution. It makes up for the shortcomings of BP neural networks, which rely solely on data-driven approaches and ignore physical decay laws. It can effectively correct the preliminary prediction results of BP neural networks.

[0040] S33: Construction of the Coupled Model: The preliminary prediction results of the improved BP neural network are coupled with the correction values ​​of the drug kinetic decay model. The final predicted value of the remaining drug lifetime is obtained by weighted summation. The output formula of the coupled model is: ;in, The final predicted remaining lifespan of the drug solution (in hours) is the final output of the model. The initial prediction value (in h) of the improved BP neural network reflects the comprehensive influence of multi-dimensional feature parameters on the lifespan of the drug solution; L(t) is the correction value (in h) of the drug solution kinetic decay model, reflecting the physical decay law of the drug solution; ω is the weighting coefficient, ranging from 0.6 to 0.8, used to adjust the influence weights of the BP neural network and the kinetic decay model. Since the BP neural network can comprehensively consider the synergistic effect of multi-dimensional feature parameters, it is given a higher weight. The value of ω is obtained by fitting the validation set data to ensure the highest prediction accuracy of the coupled model. Through the coupled model, the nonlinear fitting advantage of the improved BP neural network is leveraged to handle complex problems of multi-factor synergistic influence, while the physical laws of the kinetic decay model are combined to ensure the rationality and accuracy of the prediction results, effectively improving the accuracy of drug solution lifespan prediction.

[0041] S4. Model Training, Validation and Adaptive Correction: Based on historical data and real-time collected effective feature parameters, the prediction model is trained, validated and adaptively corrected to ensure that the model prediction accuracy meets the preset requirements, while improving the model's adaptability and stability, enabling long-term stable operation.

[0042] S41. Data Partitioning: Divide the preprocessed effective feature parameter set into a training set, validation set, and test set in a 7:2:1 ratio. The training set is used for model training, the validation set is used to adjust model parameters, and the test set is used to verify the model's prediction accuracy. For example, if the effective feature parameter set contains 1000 sets of data, then the training set will have 700 sets, the validation set 200 sets, and the test set 100 sets, ensuring sufficient data to adequately train the model.

[0043] S42. Model Training: Input the training set into the improved BP neural network, initialize model parameters including weights, threshold, learning rate, momentum factor, etc., set the number of training iterations (1000-5000) and the training error threshold (≤0.001), use momentum gradient descent to update weights, introduce dropout to prevent overfitting, and employ an adaptive learning rate adjustment strategy to dynamically adjust the learning rate until the training error reaches the preset threshold or the maximum number of training iterations is reached, completing the initial training of the BP neural network. During training, monitor changes in the training error in real time. If the training error does not decrease for 100 consecutive iterations, stop training to avoid overfitting due to overtraining.

[0044] S43. Model calibration: Initializing the predicted values ​​output by the trained BP neural network. Substituting these values ​​into the drug kinetic decay model, the correction value L(t) of the kinetic decay model is calculated. Then, based on the output formula of the coupled model, the prediction result of the coupled model is calculated. Using validation set data, the weight coefficients ω, correction coefficients α, and decay coefficients λ of the coupled model are adjusted. A grid search method is used to traverse the range of values ​​for each parameter to find the parameter combination that minimizes the prediction error on the validation set, thus completing the model calibration. (For example, the weight coefficient ω ranges from 0.6 to 0.8 with a step size of 0.05, the correction coefficient α ranges from 0.05 to 0.15 with a step size of 0.01, and the decay coefficient λ ranges from 0.02 to 0.08h.) -1 With a step size of 0.01, a grid search is used to find the parameter combination that minimizes the mean absolute error (MAE) of the validation set, which is then used as the final parameters of the coupled model.

[0045] S44. Model Validation: Input the test set into the calibrated coupled model and calculate the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) between the predicted and actual drug lifespan. 2 ),in: Mean Absolute Error (MAE): Reflects the average deviation between predicted and actual values. The formula for calculation is: ; Root Mean Square Error (RMSE): Reflects the degree of dispersion between the predicted and actual values. The calculation formula is: ; Coefficient of determination (R) 2 Ri reflects the goodness of fit of the model. 2 The closer the value is to 1, the better the model fits the data. The calculation formula is: ; Where m is the number of samples in the test set. Let be the predicted value for the i-th sample. For the actual value of the i-th sample, This is the average of the actual values. In this invention, the model's preset accuracy requirement is: R 2 If the validation results on the test set meet the above requirements (≥0.95, MAE≤0.5h, RMSE≤0.8h), the model training is complete and the model can be used. If not, return to step S42, adjust the model parameters (such as the number of hidden layer neurons, the initial value of the learning rate, the momentum factor, etc.) and retrain until the preset accuracy requirements are met.

[0046] S45. Adaptive Correction: In practical applications, changes in drug formulation, process conditions, and environmental conditions may occur, leading to a decrease in the model's prediction accuracy. Therefore, adaptive correction of the model is necessary. Specifically, dynamic characteristic parameters and actual lifespan data of the drug solution are collected in real time. Newly collected data is added to the training set periodically (every 24 hours) to fine-tune the coupled model, updating the model's weights, thresholds, weight coefficients ω, correction coefficients α, and attenuation coefficients λ. This allows the model to adapt to changes in drug composition and fluctuations in operating conditions, ensuring that the model's long-term prediction accuracy meets preset requirements. Simultaneously, a model performance monitoring mechanism is established to calculate the model's prediction error in real time. If the prediction error exceeds the preset threshold (MAE > 0.5h) for five consecutive times, the adaptive correction process is immediately triggered without waiting 24 hours, ensuring the model can adapt to sudden changes promptly.

[0047] S5. Real-time Prediction and Failure Early Warning: Dynamic characteristic parameters of the alkaline blasting cleaning solution are collected in real time, preprocessed according to the method in step S2 to obtain effective characteristic parameters, and input into the trained prediction model. The model outputs the predicted value of the remaining lifespan of the cleaning solution. It also provides failure warning signals, enabling real-time monitoring of the liquid's lifespan and early warning of failure.

[0048] The failure warning signal is divided into three levels, based on the predicted remaining lifespan of the solution, as follows: Level 1 warning: The remaining lifespan of the solution is ≥2 hours, reminding operators to pay attention to the status of the solution and strengthen the monitoring of key parameters such as effective alkali concentration and silicon ion concentration to ensure the stability of the solution performance; Level 2 warning: 1h≤remaining lifespan of the solution<2h, prompting operators to prepare to replace or replenish the solution, and to make preparations in advance to avoid production interruption due to solution failure; Level 3 warning: If the remaining lifespan of the cleaning solution is less than 1 hour, an audible and visual alarm signal will be issued immediately (audible alarm volume ≥ 80dB, visual alarm flashing red). At the same time, the warning signal will be pushed to the operator's terminal through the host computer system, prompting the operator to stop the cleaning operation and replace the cleaning solution immediately to prevent the production of defective products.

[0049] The warning signals are output in the following ways: on-site audible and visual alarms, upper computer system prompts, and mobile APP push notifications, ensuring that operators can receive warning information in a timely manner. At the same time, the warning time, warning level, and corresponding characteristic parameters are recorded to facilitate subsequent traceability and analysis.

[0050] S6. Generate suggestions for replacing or replenishing the solution: Based on the predicted remaining lifespan of the solution and real-time data on effective alkali concentration and silicon ion concentration, generate targeted suggestions for replacing or replenishing the solution to guide operators in subsequent operations and achieve intelligent management of the solution.

[0051] The specific suggestion generation rules are as follows: When the remaining lifespan of the cleaning solution is less than 1 hour, the effective alkali concentration is lower than the preset threshold (1.5%), or the silicon ion concentration is higher than the preset threshold (5000ppm), a suggestion to immediately replace the cleaning solution is generated, clearly instructing the operator to stop the cleaning operation, drain the old cleaning solution, and replace it with a new cleaning solution. At the same time, the formula and preparation method of the new cleaning solution are given (e.g., the effective alkali concentration is 3%-5%, and the surfactant concentration is 1%-2%). When 1 hour ≤ remaining lifespan of the solution < 2 hours and the effective alkali concentration is within the preset threshold range (1.5%-3%), a supplementary solution recommendation is generated. The supplementary amount is calculated based on the difference between the effective alkali concentration and the standard concentration, and the total amount of solution. The formula for calculating the supplementary amount is as follows: ; The meanings of each parameter are as follows: The volume of medicine to be replenished (unit: L); The total volume of the cleaning solution for the alkaline polishing machine (unit: L) is determined according to the model of the cleaning machine. For example, the total volume of the cleaning solution for a conventional photovoltaic silicon wafer alkaline polishing machine is 500L. The effective alkali standard concentration of the solution (unit: %), with a range of 3%-5%, is set according to process requirements; The real-time effective alkali concentration of the drug solution (unit: %) is collected by an online concentration detector. To replenish the effective alkali concentration (unit: %) of the cleaning solution, the value should be between 10% and 15%, higher than the standard concentration, to ensure that the effective alkali concentration of the cleaning solution can quickly reach the standard value after replenishment. (For example, the total volume of the cleaning solution in the alkaline polishing machine...) Effective alkali standard concentration Real-time concentration Replenish the effective alkali concentration of the medicine solution Then the amount to supplement This means that 62.5L of effective alkali solution with a concentration of 12% needs to be added.

[0052] The generated suggestions also include the timing, operating procedures, and precautions for replacing or replenishing the cleaning solution. For example, the operating procedures for replacing the cleaning solution are: stop the cleaning operation → drain the old solution → clean the solution tank → prepare the new solution → inject the solution into the solution tank → check the solution parameters → start the cleaning operation. The precautions are: wear protective equipment (gloves, goggles, protective clothing) when replacing the solution to avoid contact with the skin and eyes; mix the solution slowly according to the ratio when preparing the solution to prevent violent reaction and splashing.

[0053] S7. Data Storage and Traceability: To facilitate subsequent traceability analysis, model optimization, and process improvement, collected dynamic characteristic parameters, input and output data of the prediction model, chemical solution replacement records, and early warning records will be stored in a database for at least one year, meeting the relevant standards for production traceability and quality control in the semiconductor and photovoltaic industries. The database uses a relational database (such as MySQL), supporting data querying, statistics, and analysis. Operators can use the host computer system to query historical data by time range, equipment number, chemical solution batch, etc., to trace changes in chemical solution performance, early warning triggering reasons, and chemical solution replacement status during specific periods, and analyze the changing patterns of chemical solution lifespan, providing data support for process parameter optimization and model parameter iteration. Simultaneously, the database supports both automatic and manual backup functions. Automatic backup is performed daily at midnight, with backup data stored on an independent server and retained for at least six months. Manual backup can be performed by operators at any time according to production needs to prevent data loss and ensure data security and integrity. In addition, a data access control mechanism is established to differentiate access permissions for different roles such as administrators, operators, and technicians. Administrators have full permissions, including data modification, backup management, and permission allocation; operators can only query data related to the equipment they are responsible for and cannot modify it; technicians can query all data for analysis and optimization, ensuring data confidentiality and standardization. The data traceability process adopts a "one-click traceability" mode. By entering the product batch number or cleaning period, the corresponding time period's chemical parameters, prediction results, early warning records, and chemical replacement information can be quickly retrieved, enabling rapid location of product quality issues and providing a basis for anomaly handling and responsibility determination.

[0054] Example 2: This application provides a big data-based system for predicting the lifespan of chemicals used in alkaline polishing machines, referring to... Figure 2 A big data-based system for predicting the lifespan of alkaline polishing cleaning solutions includes the following: The data acquisition module is electrically connected to the input of the data processing module. It is used to build a multi-dimensional characteristic parameter acquisition system for the alkaline polishing cleaning machine solution. The system collects the dynamic characteristic parameters of the solution at different working times. The dynamic characteristic parameters include chemical characteristic parameters, physical characteristic parameters, and process condition parameters. Data processing module: The output end is electrically connected to the input end of the prediction module. It is used to preprocess the dynamic feature parameters collected by the acquisition module, including outlier removal, data standardization, feature filtering, removal of redundant and outlier data, elimination of the influence of units, and obtaining an effective set of feature parameters for model training. Prediction module: The output terminal is electrically connected to the input terminal of the training module and is used to build a prediction model. The prediction model is a coupled model of an improved BP neural network and a drug liquid kinetic decay model. The improved BP neural network is used to initially predict the remaining life of the drug liquid, and the kinetic decay model is used to correct the initial prediction results. Training module: The output end is electrically connected to the input end of the early warning module. Based on historical data and real-time collected effective feature parameters, it trains, validates, and adaptively corrects the prediction model. Early warning module: The output end is electrically connected to the input end of the feedback module. It is used to collect the dynamic characteristic parameters of the alkali blasting cleaning liquid in real time, preprocess it according to the method of the data processing module to obtain effective characteristic parameters, and input them into the trained prediction model. The model outputs the remaining life of the liquid and failure early warning signal. Feedback module: The output end is electrically connected to the input end of the data storage module. Based on the predicted remaining lifespan of the drug solution and real-time data of effective alkali concentration and silicon ion concentration, it generates targeted suggestions for drug solution replacement or replenishment. Data storage module: Stores dynamic feature parameters collected by the acquisition module, input and output data of the prediction model, drug solution replacement records, early warning records, and other data in the database.

[0055] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the lifespan of alkaline polishing cleaning solutions based on big data, characterized in that, Includes the following steps: S1. Establish a multi-dimensional feature parameter acquisition system: Establish a multi-dimensional feature parameter acquisition system for the alkaline polishing cleaning machine solution to collect dynamic feature parameters of the solution at different working periods. The dynamic feature parameters include chemical feature parameters, physical feature parameters and process condition parameters. S2: Feature parameter preprocessing: The collected dynamic feature parameters are preprocessed, including outlier removal, data standardization, feature selection, removal of redundant and outlier data, elimination of the influence of units, and obtaining an effective set of feature parameters for model training. S3: Constructing a drug liquid life prediction model: The prediction model is a coupled model of an improved BP neural network and a drug liquid kinetic decay model, wherein the improved BP neural network is used to initially predict the remaining life of the drug liquid, and the kinetic decay model is used to correct the initial prediction results; S4: Model Training, Validation and Adaptive Correction: Based on historical data and effective feature parameters acquired in real time, the prediction model is trained, validated and adaptively corrected. S5: Real-time prediction and failure warning: Real-time acquisition of dynamic characteristic parameters of the alkaline blasting cleaning liquid, preprocessing according to the method of step S2 to obtain effective characteristic parameters, inputting them into the trained prediction model, and the model outputs the predicted value of the remaining life of the liquid Lfinal and the failure warning signal. S6: Generate suggestions for replacing or replenishing the solution: Based on the predicted remaining lifespan of the solution and real-time data on effective alkali concentration and silicon ion concentration, generate targeted suggestions for replacing or replenishing the solution.

2. The method for predicting the lifespan of alkaline polishing cleaning solutions based on big data according to claim 1, characterized in that, In step S1, the chemical characteristic parameters include: effective alkali concentration, silicon ion concentration, impurity ion concentration, surfactant concentration, and pH value; wherein the effective alkali concentration is the mass concentration of sodium hydroxide or potassium hydroxide, the silicon ion concentration is the concentration of silicate ions generated by the reaction of the solution with the silicon wafer, and the impurity ion concentration includes the concentrations of iron ions, copper ions, and sodium ions; the physical characteristic parameters include the temperature, viscosity, density, and surface tension of the solution; and the process condition parameters include the material of the workpiece being cleaned, the number of workpiece batches, the single cleaning time, the set value of the cleaning temperature, the circulation speed of the solution, and the frequency and amount of replenishment of the solution.

3. The method for predicting the lifespan of alkaline polishing cleaning solutions based on big data according to claim 1, characterized in that, The specific process of step S2 is as follows: S21. Outlier Removal: Using the Grubbs criterion, outlier detection is performed on the collected dynamic feature parameters, and outlier data exceeding 3 times the standard deviation are removed. S22. Data Standardization: The min-max standardization method is used to map the feature parameters after removing outliers to the [0, 1] interval, eliminating the influence of different dimensions of parameters. The standardization formula is: ; Here, x represents the standardized feature parameter values, and x represents the original feature parameter values. This is the minimum value of the characteristic parameter. This is the maximum value of the feature parameter; S23. Feature Screening: Using the mutual information method, the mutual information value between each feature parameter and the lifespan of the liquid is calculated. Feature parameters with mutual information values ​​greater than the preset threshold (0, 3) are selected to form an effective feature parameter set, and redundant parameters with weak correlation to the lifespan of the liquid are removed.

4. The method for predicting the lifespan of alkaline polishing cleaning solutions based on big data according to claim 1, characterized in that, In step S3, the improved BP neural network structure includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is the same as the number of parameters in the effective feature parameter set. The hidden layer has 2-3 layers, with each layer containing 1.5-2 times the number of neurons in the input layer. The output layer has only one neuron, and its output is the predicted remaining lifespan of the drug solution. The improved BP neural network uses momentum gradient descent to optimize weight updates, introduces a dropout mechanism to prevent overfitting, and employs an adaptive learning rate adjustment strategy, dynamically adjusting the learning rate based on the training error. The learning rate adjustment formula is: ； in, Let be the learning rate for the (k+1)th iteration. Let be the learning rate for the k-th iteration. Let be the training error of the k-th iteration. This represents the training error of the (k+1)th iteration; The drug solution kinetic decay model is constructed based on the consumption law of the active ingredient in the drug solution, and its expression is: ； Where L(t) is the remaining lifespan of the drug solution at time t. Let λ be the initial lifetime of the drug solution, λ be the decay coefficient of the effective components of the drug solution, t be the working time of the drug solution, α be the correction coefficient (range 0.05-0.15), and n be the number of impurity ions. Let be the concentration of the i-th impurity ion at time t. The weight of the influence of the i-th impurity ion on the lifespan of the drug solution; The output formula of the coupled model is: ； in, This is the final predicted value for the remaining lifespan of the drug solution. The values ​​are the initial predictions of the improved BP neural network, where ω is the weight coefficient (ranging from 0.6 to 0.8) and L(t) is the correction value of the dynamic decay model.

5. The method for predicting the lifespan of alkaline polishing cleaning solutions based on big data according to claim 1, characterized in that, The specific process of step S4 is as follows: S41. Data partitioning: Divide the effective feature parameter set into training set, validation set and test set in a ratio of 7:2:

1. The training set is used for model training, the validation set is used to adjust model parameters and the test set is used to verify the model prediction accuracy. S42. Model Training: Input the training set into the improved BP neural network, initialize the model parameters (weights, threshold, learning rate, momentum factor), set the number of training iterations (1000-5000 times) and the training error threshold (≤0.001), and use the momentum gradient descent method to update the weights until the training error reaches the preset threshold or the maximum number of training iterations is reached, thus completing the initial training of the BP neural network. S43. Model calibration: Substitute the initial predicted values ​​output by the trained BP neural network into the drug kinetic decay model for calibration to obtain the prediction results of the coupled model. Use the validation set to adjust the weight coefficient ω, correction coefficient α and decay coefficient λ of the coupled model to minimize the prediction error of the validation set. S44. Model Validation: Input the test set into the calibrated coupled model and calculate the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) between the predicted and actual drug lifespan. 2 If R 2 If the accuracy is ≥0.95, MAE≤0.5h, and RMSE≤0.8h, then the model meets the preset accuracy requirements; if not, return to step S42, adjust the model parameters, and retrain. S45. Adaptive correction: Real-time acquisition of dynamic characteristic parameters and actual lifespan data of the liquid, periodic fine-tuning of the coupled model, updating model parameters, and adapting to changes in liquid composition and fluctuations in operating conditions.

6. The method for predicting the lifespan of alkaline polishing cleaning solutions based on big data according to claim 1, characterized in that, In step S5, the failure warning signal is divided into three levels: Level 1 warning (remaining lifespan of the cleaning solution ≥ 2h), prompting the operator to pay attention to the status of the cleaning solution and strengthen parameter monitoring; Level 2 warning (1h ≤ remaining lifespan of the cleaning solution < 2h), prompting the operator to prepare to replace or replenish the cleaning solution; Level 3 warning (remaining lifespan of the cleaning solution < 1h), immediately issuing an alarm signal, prompting the operator to stop the cleaning operation and replace the cleaning solution; the warning signal is output synchronously through the audible and visual alarm device and the host computer system, and the warning time and corresponding characteristic parameters are recorded at the same time; In step S6, the suggestion to replace or replenish the solution is generated based on the predicted remaining lifespan of the solution and real-time data of the effective alkali concentration and silicon ion concentration: when the remaining lifespan of the solution is <1h or the effective alkali concentration is lower than the preset threshold (1.5%) or the silicon ion concentration is higher than the preset threshold (5000ppm), a suggestion to replace the solution immediately is generated; when 1h ≤ remaining lifespan of the solution <2h and the effective alkali concentration is within the preset threshold range (1.5%-3%), a suggestion to replenish the solution is generated. The replenishment amount is calculated based on the difference between the effective alkali concentration and the standard concentration and the total amount of solution. The formula for calculating the replenishment amount is as follows: ； in, This is to replenish the amount of medicine. This refers to the total volume of the cleaning solution in the alkaline polishing machine. The effective alkali standard concentration of the drug solution is 3%-5%. This represents the real-time effective alkali concentration of the drug solution. To replenish the effective alkali concentration of the solution (10%-15%), the generated suggestions also include the time points for solution replacement or replenishment, operating procedures and precautions, which are simultaneously pushed to the operator's terminal.

7. The method for predicting the lifespan of alkaline polishing cleaning solutions based on big data according to claim 1, characterized in that, In step S1, the characteristic parameters are collected every 5-10 minutes. The effective alkali concentration is collected by titration combined with an online concentration detector, the silicon ion concentration is collected by spectrophotometry, the impurity ion concentration is collected by inductively coupled plasma optical emission spectrometry (ICP-OES), the pH value is collected by an online pH meter, and the temperature, viscosity, density, and surface tension are collected by online temperature sensors, viscosity sensors, density sensors, and surface tension sensors, respectively. The process parameters are read in real time by the PLC control system of the alkali polishing cleaning machine.

8. The method for predicting the lifespan of alkaline polishing cleaning solutions based on big data according to claim 4, characterized in that, It also includes step S7, data storage and traceability: storing the collected dynamic feature parameters, the input and output data of the prediction model, the drug solution replacement record, the early warning record and other data into the database.

9. A big data-based system for predicting the lifespan of chemicals used in alkaline blasting cleaning machines, employing the big data-based method for predicting the lifespan of chemicals used in alkaline blasting cleaning machines as described in any one of claims 1-8, characterized in that: include: The data acquisition module is electrically connected to the input of the data processing module. It is used to build a multi-dimensional characteristic parameter acquisition system for the alkaline polishing cleaning machine solution. The system collects the dynamic characteristic parameters of the solution at different working times. The dynamic characteristic parameters include chemical characteristic parameters, physical characteristic parameters, and process condition parameters. Data processing module: The output end is electrically connected to the input end of the prediction module. It is used to preprocess the dynamic feature parameters collected by the acquisition module, including outlier removal, data standardization, feature filtering, removal of redundant and outlier data, elimination of the influence of units, and obtaining an effective set of feature parameters for model training. Prediction module: The output terminal is electrically connected to the input terminal of the training module and is used to build a prediction model. The prediction model is a coupled model of an improved BP neural network and a drug liquid kinetic decay model. The improved BP neural network is used to initially predict the remaining life of the drug liquid, and the kinetic decay model is used to correct the initial prediction results. Training module: The output end is electrically connected to the input end of the early warning module. Based on historical data and real-time collected effective feature parameters, it trains, validates, and adaptively corrects the prediction model. Early warning module: The output end is electrically connected to the input end of the feedback module. It is used to collect the dynamic characteristic parameters of the alkali blasting cleaning liquid in real time, preprocess it according to the method of the data processing module to obtain effective characteristic parameters, and input them into the trained prediction model. The model outputs the remaining life of the liquid and failure early warning signal. Feedback module: The output end is electrically connected to the input end of the data storage module. Based on the predicted remaining lifespan of the drug solution and real-time data of effective alkali concentration and silicon ion concentration, it generates targeted suggestions for drug solution replacement or replenishment. Data storage module: Stores dynamic feature parameters collected by the acquisition module, input and output data of the prediction model, drug solution replacement records, early warning records, and other data in the database.

Images