Method and system for evaluating filtering effect of semiconductor preparation waste gas

Abstract

The invention discloses a semiconductor preparation waste gas filtering effect evaluation method and system, and particularly relates to the technical field of waste gas filtering. The method comprises the following steps: synchronously acquiring response data through multiple types of sensors, extracting interference characteristics including abnormal spectrogram morphological characteristics and time change gradient characteristics, constructing a machine learning model to predict error severity based on historical experimental data and a gas interference mechanism, and implementing a differential concentration correction strategy according to an error level, so as to improve the accuracy of concentration correction. According to the method, intelligent recognition and correction of cross interference are achieved, and the accuracy of waste gas monitoring and the reliability of system evaluation are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of waste gas filtration, and in particular to a method and system for evaluating the filtering effect of semiconductor production waste gas. Background Art

[0002] Semiconductor manufacturing waste gas filtration effect evaluation refers to the systematic analysis and evaluation of the purification efficiency and treatment effect of the filtration system through a series of methods and indicators after the waste gas generated in the semiconductor manufacturing process is filtered and treated. This evaluation usually includes the detection and analysis of harmful gas removal rate, filter material performance, and emission gas concentration changes, with the aim of ensuring that waste gas emissions meet environmental protection standards, ensuring production safety and reducing the impact on the environment.

[0003] The prior art has the following deficiencies: In the semiconductor manufacturing process, the evaluation of the exhaust gas filtration effect is crucial. However, in the existing technology, when the continuous emission monitoring system (CEMS) detects the target gas, it often produces cross-interference errors due to interference from other components, resulting in inaccurate readings. For example, NDIR or electrochemical sensors may Mistakenly judged as , or because Existence and interference This can cause errors in pollutant concentration or purification efficiency assessment. Such errors usually occur when multiple gases coexist, their concentrations are close, or the instrument is not calibrated strictly, which poses a challenge to high-precision assessment. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for evaluating the filtering effect of semiconductor manufacturing waste gas to solve the shortcomings of the background technology.

[0005] In order to achieve the above object, the present invention provides the following technical solution: a method for evaluating the filtering effect of semiconductor manufacturing waste gas, comprising: Collect response data obtained by multiple gas sensors of different types performing synchronous detection on the same exhaust gas to be tested; Extracting gas interference features from the response data of the different sensors, the interference features including abnormal spectrum morphology features and time-varying gradient features between sensors; Based on historical experimental data and known interference gas impact mechanisms, a machine learning algorithm is used to train a cross-interference prediction model, and the gas interference characteristics collected in real time are input into the cross-interference prediction model to predict the severity of errors caused by cross-interference in the current monitoring data; According to the severity of the error, different gas concentration correction strategies are selected to perform differential correction on the original monitoring data of the target gas to obtain the target gas concentration value after cross-interference correction; Based on the corrected target gas concentration value, the purification efficiency of the semiconductor exhaust gas filtration system is evaluated.

[0006] Preferably, the different sensors include NDIR infrared sensors, electrochemical gas sensors, Fourier transform infrared spectrometers and mass spectrometers.

[0007] Preferably, the interference feature includes abnormal spectral morphological features between sensors, and the extraction method is: two sensors A and B are provided, and n time points are sampled in the same time window to obtain two response sequences respectively: ;in: is the original response value of sensor A at the nth time point, is the original response value of sensor B at the nth time point. The two sequences are normalized respectively. After normalization, the Euclidean distance between the two sequences is calculated as the morphological feature of the abnormal spectrum.

[0008] Preferably, the interference feature includes a time-varying gradient feature between sensors, and the extraction method is: assuming that the discrete signal sequence of the sensor response that varies with time is: ; The corresponding time points are: ;in represents the sensor response value at the mth time point, For the corresponding timestamp, the timestamp As the horizontal axis, the sensor response value As the vertical axis, a rectangular coordinate system is constructed to generate the sensor response curve S(t) and calculate its first-order derivative ; Select a sliding time window w of fixed length, and in each window , the difference between the maximum rate of change and the minimum rate of change of the sensor response is calculated as the time-varying gradient feature.

[0009] Preferably, the abnormal spectrum morphological features and time-varying gradient features are converted into comprehensive feature vectors, and the comprehensive feature vectors are used as inputs of the machine learning model. The machine learning model uses each group of comprehensive feature vectors to predict the error severity value label caused by cross-interference in the current monitoring data as a prediction target, and takes minimizing the sum of prediction errors of the error severity value labels caused by cross-interference in all current monitoring data as a training target. The machine learning model is trained until the sum of prediction errors converges, and the model training is stopped. The error severity value caused by cross-interference in the current monitoring data is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

[0010] Preferably, the error severity value caused by the cross interference in the acquired current monitoring data is compared with the gradient standard threshold, the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and the error severity value caused by the cross interference in the current monitoring data is compared with the first standard threshold and the second standard threshold respectively; If the error severity value is greater than the second standard threshold, the current monitoring data is judged to be in a high-risk interference state, and the alarm mechanism must be triggered immediately or the redundant sensor data correction process must be started; if the error severity value is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it is judged to be a medium-risk interference state; if the error severity value is less than the first standard threshold, it is judged to be a low-risk interference state.

[0011] Preferably, the original concentration value Craw of the target gas at each moment is corrected according to the selected strategy, and the corrected concentration value Ccorrected is calculated: ; Where ΔC is the concentration error value caused by cross interference, which is defined as the difference between the sensor output and the actual target gas concentration; The corrected target gas concentration Ccorrected is output as the true concentration estimate at the current time point.

[0012] Preferably, the purification efficiency calculation formula of the semiconductor exhaust gas filtration system is: ; Where η is the purification efficiency, is the target gas concentration before the exhaust gas enters the filtration system, is the target gas concentration after the exhaust gas is treated; For each set of sampling data, the purification efficiency value is calculated based on the corrected inlet and outlet concentrations; The purification efficiency results are displayed visually. If the efficiency is lower than the set threshold, an early warning prompt is triggered to assist maintenance personnel in making judgments and optimization adjustments.

[0013] The present invention also provides a semiconductor manufacturing waste gas filtering effect evaluation system, including a data acquisition module, a feature extraction module, an error prediction module, a correction module and a purification efficiency evaluation module; Data acquisition module: collects the response data obtained by multiple gas sensors of different types for synchronous detection of the same exhaust gas to be tested; Feature extraction module: extracting gas interference features from the response data of the different sensors, wherein the interference features include abnormal spectrum morphology features and time-varying gradient features between sensors; Error prediction module: Based on historical experimental data and known interference gas influence mechanisms, a machine learning algorithm is used to train a cross-interference prediction model, and the real-time collected gas interference features are input into the cross-interference prediction model to predict the severity of the error caused by cross-interference in the current monitoring data; Correction module: according to the severity of the error, different gas concentration correction strategies are selected to perform differential correction on the original monitoring data of the target gas to obtain the target gas concentration value after cross-interference correction; Purification efficiency evaluation module: Based on the corrected target gas concentration value, the purification efficiency of the semiconductor exhaust gas filtration system is evaluated.

[0014] In the above technical solution, the technical effects and advantages provided by the present invention are: 1. The present invention extracts representative abnormal spectral morphological features and time-varying gradient features by fusing multi-type sensor response data, uses machine learning models to intelligently predict the severity of errors caused by cross-interference, and implements differentiated concentration correction strategies based on different interference levels, thereby effectively improving the accuracy of target gas concentration data. This method overcomes the recognition error problem caused by the coexistence of multiple gases and poor sensor selectivity in traditional continuous emission monitoring systems, and is particularly suitable for high-precision dynamic monitoring of complex mixed exhaust gases in semiconductor manufacturing processes.

[0015] 2. The present invention constructs a purification efficiency evaluation model based on the corrected target gas concentration value, which can realize a scientific, real-time and visual evaluation of the purification effect of the semiconductor exhaust gas filtration system. When the purification efficiency is lower than the set threshold, the system can trigger an early warning in time, which helps the operation and maintenance personnel to find problems such as aging of the filter material or system abnormalities, and realize closed-loop control of the whole process. The present invention has significantly improved the monitoring reliability, evaluation accuracy and intelligence level of the exhaust gas treatment system as a whole, and has good engineering practical value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0017] Figure 1 The figure is a flow chart of the method of the present invention.

[0018] Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] Example 1, please refer to Figure 1 As shown, the semiconductor manufacturing waste gas filtering effect evaluation method described in this embodiment includes: Collect response data obtained by multiple gas sensors of different types performing synchronous detection on the same exhaust gas to be tested; Extracting gas interference features from the response data of the different sensors, the interference features including abnormal spectrum morphology features and time-varying gradient features between sensors; Based on historical experimental data and known interference gas impact mechanisms, a machine learning algorithm is used to train a cross-interference prediction model, and the gas interference characteristics collected in real time are input into the cross-interference prediction model to predict the severity of errors caused by cross-interference in the current monitoring data; According to the severity of the error, different gas concentration correction strategies are selected to perform differential correction on the original monitoring data of the target gas to obtain the target gas concentration value after cross-interference correction; Based on the corrected target gas concentration value, the purification efficiency of the semiconductor exhaust gas filtration system is evaluated.

[0021] In a complex semiconductor exhaust environment, different gases have different chemical properties, molecular structures, spectral responses and other characteristics, so a single type of sensor is often difficult to fully and accurately perceive. For this reason, the system is usually equipped with the following types of sensors: NDIR infrared sensor (Non-Dispersive Infrared): Mainly used to detect gases with infrared absorption characteristics, such as , , This type of sensor measures the absorption intensity of infrared light of a specific wavelength to infer the gas concentration. It has a fast response and a long life, but it is easily interfered and its accuracy is reduced when multiple gases with similar absorption bands coexist.

[0022] Electrochemical gas sensors: suitable for detecting , , , The gas to be detected generates a current signal through an electrochemical reaction on the electrode surface, thereby calculating the concentration. Although it has high sensitivity and low power consumption, it has problems such as poor selectivity and sensitivity to temperature and humidity.

[0023] Fourier Transform Infrared Spectrometer (FTIR): It can identify and quantitatively analyze multiple gas components simultaneously. This technology is based on the absorption characteristics of molecules in the mid-infrared region and can provide a complete absorption spectrum. Although it is large in size and high in cost, it has strong resolution and anti-interference capabilities and is an important tool for complex gas mixture analysis.

[0024] Mass spectrometer (such as GC-MS): It can realize qualitative and quantitative detection of trace components and is suitable for low-concentration, high-precision analysis, especially for monitoring and tracing trace impurities. It is usually used for offline verification or calibration of models, and is not often used in real-time online, but its data is crucial for model training.

[0025] In actual deployment, these sensors are placed at the inlet and outlet of the exhaust gas filtration unit to achieve real-time monitoring of untreated and treated gases. The system uses a unified data acquisition module (such as an industrial Internet of Things acquisition card or an edge computing terminal) to synchronously trigger various sensors for data acquisition with millisecond-level accuracy, ensuring that the data of all sensors can correspond to the state of the exhaust gas at the same time, avoiding misjudgment caused by time differences.

[0026] The collected data includes raw signals (such as absorbed light intensity, current, voltage, spectrum, etc.). The sampling frequency of each type of sensor is set according to its response speed. For example, NDIR sensors can reach 1 Hz, electrochemical sensors are usually 0.1 to 1 Hz, and FTIR equipment is usually scanned several times per second.

[0027] The multi-dimensional data collected by various sensors are not only complementary in terms of accuracy, but more importantly, they enhance each other in terms of information dimension: when the NDIR sensor makes fuzzy judgments due to the overlap of certain gas spectra, it can be corrected by the full spectrum absorption data of FTIR; electrochemical sensors are sensitive to certain small molecule gases and can be used as auxiliary judgment basis for NDIR systems; multiple signals perform "redundant verification" on the same gas, which significantly improves the confidence of judgment; when a sensor has a short-term drift or data anomaly, the system can detect anomalies and automatically replace compensation by comparing the outputs of other sensors to enhance overall robustness. In addition, the system can also establish a cross-reference model or joint feature vector based on different sensor data to provide more comprehensive and accurate data support for subsequent gas concentration inversion models or cross-interference correction models.

[0028] In this application, by deploying multiple types of sensors and performing synchronous detection, not only can the recognition accuracy of complex exhaust gas components be improved, but the risk of misjudgment caused by cross-interference can also be effectively dealt with. This technical solution lays a solid foundation for building a highly reliable and intelligent exhaust gas filtration effect evaluation system, and also provides a rich and reliable data source for subsequent algorithm modeling and model correction. It reflects the current development trend of exhaust gas detection systems from "single-point perception" to "multi-source fusion intelligent identification."

[0029] When the sensor responds to the target gas, it will output a curve or spectrum of an electrical signal (such as voltage, current, infrared absorption intensity, etc.) that changes over time. If the sensor suffers from cross interference, its response curve will show abnormal morphological characteristics, which are different from the characteristic profile under the standard state. The so-called abnormal spectrum morphological characteristics refer to the inconsistent or unreasonable deviation in shape, position, amplitude, etc. of the response curves of multiple sensors to the same exhaust gas in the same time window. By extracting these anomalies, it can be determined whether there is cross interference.

[0030] The following methods can be used to extract and quantify spectral features: Waveform comparison differences: After normalizing the output curves of two or more sensors, calculate the Euclidean distance, dynamic time warping (DTW) distance, or similarity measure (such as cosine similarity) between them; Peak drift detection: Analyzes whether the peak time position or peak amplitude in each sensor response curve is offset. Typical features include peak shift, waveform misalignment, abnormal double peak structure, etc. Spectral profile anomaly detection: Use Fourier transform or wavelet transform to extract frequency domain features and compare the spectrum distribution of different sensors. If there is inconsistent spectral energy distribution, missing frequency bands or multi-frequency anomalies, it is regarded as a morphological abnormality feature.

[0031] High-order statistical analysis: Extract the first-order derivative (slope), second-order derivative (curvature), skewness, kurtosis and other statistics of the response curve to compare whether there are significant differences between different sensors.

[0032] The abnormal spectral morphological features can reveal the structural differences in the response behaviors of multiple sensors under non-ideal conditions. They are highly reliable feature bases for determining whether there is non-target gas interference, and provide highly separable input variables for machine learning modeling using spectral features.

[0033] For example, the present invention can use the following method to extract abnormal spectrum morphological features: Suppose there are two sensors A and B, sampling n time points in the same time window, and obtaining two response sequences respectively: ;in: is the original response value of sensor A at the nth time point, is the original response value of sensor B at the nth time point. In order to eliminate the deviation caused by the amplitude difference, the two sequences are normalized (such as Z-score normalization or Min-Max normalization), Z-score normalization (standardization): ;in: is the mean of sequence A, is the standard deviation of sequence A. Similarly, , is the mean and standard deviation of B; , is the normalized response value.

[0034] After normalization, calculate the Euclidean distance between the two sequences , as the abnormal spectrum morphological feature, represents the difference between the two response curves, and the expression is: ; n is the total number of time points.

[0035] like ≈0, indicating that the two curves are highly consistent and there is no obvious interference in the sensor response; if If the value of the sensor response curve increases significantly, it means that there are significant differences between the sensor response curves, which may be caused by differences in target gas concentration, cross interference or failure of a sensor. The system can set an empirical threshold δ. >δ, triggering the interference detection mechanism or starting the correction model.

[0036] The time-varying gradient characteristic refers to the rate of change or trend of the sensor response value over time during continuous monitoring. If the exhaust gas components are pure and stable, the responses of each sensor should change slowly over time; but when an interfering gas suddenly enters or the concentration changes suddenly, the sensor output will change dramatically or nonlinearly. This gradient of change can reflect the instantaneous characteristics of cross-interference.

[0037] Extraction methods include: First-order derivative (Δ value) analysis: Calculate the first-order derivative of the sensor response curve S(t), that is, the amplitude of the signal change per unit time. Sudden changes or abnormal rises / falls indicate that there may be unexpected gas disturbances.

[0038] Second-order derivative (acceleration) analysis: If the first-order derivative changes dramatically, the second-order derivative is calculated to identify the turning point of the signal change. It is especially suitable for detecting response rate reversal or nonlinear abrupt increase caused by cross-interference.

[0039] Sliding window statistical features: within a fixed time window (such as 5 seconds), extract the mean, standard deviation, change amplitude, maximum jump rate, etc. for feature input of the machine learning model.

[0040] Abnormal gradient threshold recognition: Set an empirical threshold. If the gradient of change in a certain period of time exceeds the threshold, it is determined to be a potential interference state.

[0041] The time-varying gradient feature reflects the instantaneous impact of the dynamic fluctuation of exhaust gas components on the sensor response. It is suitable for identifying cross-errors caused by transient interference (such as gas pulses, valve switching and other operations). It is easy to calculate in real time, suitable for deployment in embedded edge processing units, and supports rapid early warning.

[0042] For example, the present invention extracts the time-varying gradient features in the following manner: Assume that the discrete signal sequence of the sensor response changing with time is: ; The corresponding time points are: ;in represents the sensor response value at the mth time point, is the corresponding timestamp (in seconds), usually sampled at equal intervals. As the horizontal axis, the sensor response value As the vertical axis, a rectangular coordinate system is constructed to generate the sensor response curve S(t), and its first-order derivative is calculated. The expression is: ; represents the rate of change of the sensor response in the ith time interval, represents the sensor response value at the i-th time point, is the corresponding timestamp; Select a sliding time window of fixed length w (such as 5 seconds, corresponding to w sampling points), and in each window , calculate the difference between the maximum change rate and the minimum change rate of the sensor response as the time-varying gradient feature. The larger the time-varying gradient feature, the more significant the potential cross-interference situation.

[0043] The abnormal spectrum morphological features and time-varying gradient features are converted into comprehensive feature vectors, and the comprehensive feature vectors are used as inputs of the machine learning model. The machine learning model uses each group of comprehensive feature vectors to predict the error severity value label caused by cross-interference in the current monitoring data as the prediction target, and takes minimizing the sum of prediction errors of the error severity value labels caused by cross-interference in all current monitoring data as the training target. The machine learning model is trained until the sum of prediction errors converges, and the model training is stopped. The error severity value caused by cross-interference in the current monitoring data is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

[0044] Compare the error severity value caused by cross interference in the acquired current monitoring data with the gradient standard threshold, the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, the first standard threshold (lower, set as a slight error tolerance range); the second standard threshold (higher, set as a boundary of unacceptable error); compare the error severity value caused by cross interference in the current monitoring data with the first standard threshold and the second standard threshold respectively; If the error severity value is greater than the second standard threshold, it is judged that the current monitoring data is in a high-risk interference state, and an alarm mechanism needs to be triggered immediately or a redundant sensor data correction process needs to be started.

[0045] If the error severity value is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it is judged as a medium risk interference state. The system records the state and can selectively perform soft correction or dynamic filtering to prompt the user to pay attention.

[0046] If the error severity value is less than the first standard threshold, it is judged to be a low risk or no significant interference state, and the current monitoring data can be deemed to be credible.

[0047] After obtaining the error severity value caused by cross-interference in the current monitoring data, this method further selects the corresponding gas concentration correction strategy based on the prediction result, and performs differential processing on the original monitoring data of the target gas, thereby obtaining the target gas concentration value after interference correction. The specific steps are as follows: Multiple correction strategies are preset, each corresponding to a level of interference (high, medium, low). Each strategy uses different processing methods and model accuracy to match different degrees of error risk to ensure a balance between performance and accuracy: Low error level (error severity value is less than the first standard threshold); Mild correction strategy, you can choose to use a fixed proportion correction factor or no correction to avoid excessive intervention; Medium error level (between the first and second criterion thresholds); medium complexity correction strategies, using empirical or parameterized functions for adjustment, such as local linear correction or bivariate interpolation; High error level (error severity value is greater than the second standard threshold); Strong correction strategy, calling the trained machine learning regression model or finding similar patterns based on historical data for fitting correction.

[0048] The original concentration value Craw of the target gas at each moment is corrected according to the selected strategy, and the corrected concentration value Ccorrected is calculated, which is expressed as follows: General expression: ; Where ΔC is the concentration error caused by cross-interference, defined as the difference between the sensor output and the actual target gas concentration: Example 1: Fixed ratio correction (mild correction): ; Among them, α is the small interference ratio determined empirically (such as 1%~2%).

[0049] Example 2: Correction based on model regression output (strong correction): ; where fmodel(X) is the error regression prediction value with the current feature vector as input.

[0050] The differentiated correction strategy has adaptive capabilities and can dynamically adjust correction parameters according to statistical data during system operation, such as the size of the correction factor, classification threshold, strategy calling order, etc., to adapt to the influence of sensor aging, gas composition changes, environmental disturbances, etc.

[0051] The corrected target gas concentration Ccorrected is output as the true concentration estimate at the current time point for use in exhaust gas treatment effect evaluation, emission compliance judgment and subsequent control strategies.

[0052] The purification efficiency of semiconductor exhaust gas filtration system usually refers to the removal ratio of specific target gas by the processing unit, reflecting the purification ability of the system for harmful or polluting components between intake and exhaust. The calculation formula is: ; Wherein, η is the purification efficiency (%), is the target gas concentration before the exhaust gas enters the filtration system (after interference correction), is the target gas concentration after exhaust gas treatment (also after interference correction).

[0053] Sensor systems are arranged at the air inlet and outlet of the filtration unit to monitor the target gas concentration in real time, and the aforementioned cross-interference correction algorithm is applied to all raw data to ensure that the data is accurate and reliable.

[0054] For each set of sampling data, the purification efficiency value is calculated according to the above formula based on the corrected inlet and outlet concentrations, which can realize dynamic purification efficiency evaluation on a continuous time series.

[0055] If multiple gas components and different operating conditions (flow rate, temperature and humidity, etc.) are combined, a multivariable purification efficiency model can be constructed to analyze and predict the changing trend of efficiency with system status.

[0056] The system can visualize the purification efficiency results. If the efficiency is lower than the set threshold (such as 85%), an early warning prompt will be triggered, indicating that there may be problems such as filter material saturation, equipment aging, and process out of control, assisting maintenance personnel in making judgments and optimizing adjustments.

[0057] Example 2, please refer to Figure 2 As shown, a semiconductor manufacturing waste gas filtering effect evaluation system described in this embodiment includes a data acquisition module, a feature extraction module, an error prediction module, a correction module and a purification efficiency evaluation module; Data acquisition module: collects the response data obtained by multiple gas sensors of different types for synchronous detection of the same exhaust gas to be tested; Feature extraction module: extracting gas interference features from the response data of the different sensors, wherein the interference features include abnormal spectrum morphology features and time-varying gradient features between sensors; Error prediction module: Based on historical experimental data and known interference gas influence mechanisms, a machine learning algorithm is used to train a cross-interference prediction model, and the real-time collected gas interference features are input into the cross-interference prediction model to predict the severity of the error caused by cross-interference in the current monitoring data; Correction module: according to the severity of the error, different gas concentration correction strategies are selected to perform differential correction on the original monitoring data of the target gas to obtain the target gas concentration value after cross-interference correction; Purification efficiency evaluation module: Based on the corrected target gas concentration value, the purification efficiency of the semiconductor exhaust gas filtration system is evaluated.

[0058] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0059] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding.

[0060] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example 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 performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0061] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A method for evaluating the filtering effect of semiconductor manufacturing waste gas, characterized in that: include: Collect response data obtained by multiple gas sensors of different types performing synchronous detection on the same exhaust gas to be tested; Extract gas interference features from the response data of different sensors. The interference features include abnormal spectral morphological features and time-varying gradient features between sensors. Based on historical experimental data and known interference gas impact mechanisms, a machine learning algorithm is used to train a cross-interference prediction model. The real-time collected gas interference features are input into the cross-interference prediction model to predict the severity of errors caused by cross-interference in the current monitoring data. According to the severity of the error, different gas concentration correction strategies are selected to perform differential correction on the original monitoring data of the target gas to obtain the target gas concentration value after cross-interference correction; Based on the corrected target gas concentration value, the purification efficiency of the semiconductor exhaust gas filtration system is evaluated.

2. The method for evaluating the filtering effect of semiconductor manufacturing waste gas according to claim 1, characterized in that: The different sensors include NDIR infrared sensors, electrochemical gas sensors, Fourier transform infrared spectrometers and mass spectrometers.

3. The method for evaluating the filtering effect of semiconductor manufacturing waste gas according to claim 1, characterized in that: The interference features include abnormal spectrogram morphological features between sensors, and the extraction method is as follows: two sensors A and B are provided, and n time points are sampled in the same time window, and two response sequences are obtained respectively: ;in: is the original response value of sensor A at the nth time point, is the original response value of sensor B at the nth time point. The two sequences are normalized respectively. After normalization, the Euclidean distance between the two sequences is calculated as the morphological feature of the abnormal spectrum.

4. The method for evaluating the filtering effect of semiconductor manufacturing waste gas according to claim 3, characterized in that: The interference feature includes the time-varying gradient feature between sensors, and the extraction method is: suppose the discrete signal sequence of the sensor response changing with time is: ; The corresponding time points are: ;in represents the sensor response value at the mth time point, For the corresponding timestamp, the timestamp As the horizontal axis, the sensor response value As the vertical axis, a rectangular coordinate system is constructed to generate the sensor response curve S(t) and calculate its first-order derivative ; Select a sliding time window w of fixed length and , the difference between the maximum rate of change and the minimum rate of change of the sensor response is calculated as the time-varying gradient feature.

5. The method for evaluating the filtering effect of semiconductor manufacturing waste gas according to claim 4, characterized in that: The abnormal spectrum morphological features and time-varying gradient features are converted into comprehensive feature vectors, and the comprehensive feature vectors are used as inputs of the machine learning model. The machine learning model uses each group of comprehensive feature vectors to predict the error severity value label caused by cross-interference in the current monitoring data as the prediction target, and takes minimizing the sum of prediction errors of the error severity value labels caused by cross-interference in all current monitoring data as the training target. The machine learning model is trained until the sum of prediction errors converges, and the model training is stopped. The error severity value caused by cross-interference in the current monitoring data is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

6. A method for evaluating the filtering effect of semiconductor manufacturing waste gas according to claim 5, characterized in that: Compare the error severity value caused by cross interference in the acquired current monitoring data with the gradient standard threshold, the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and compare the error severity value caused by cross interference in the current monitoring data with the first standard threshold and the second standard threshold respectively; If the error severity value is greater than the second standard threshold, it is judged that the current monitoring data is in a high-risk interference state, and an alarm mechanism needs to be triggered immediately or a redundant sensor data correction process needs to be started; If the error severity value is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it is judged as a medium risk interference state; if the error severity value is less than the first standard threshold, it is judged as a low risk interference state.

7. A method for evaluating the filtering effect of semiconductor manufacturing waste gas according to claim 6, characterized in that: The original concentration value Craw of the target gas at each moment is corrected according to the selected strategy, and the corrected concentration value Ccorrected is calculated: ; Where ΔC is the concentration error value caused by cross interference, which is defined as the difference between the sensor output and the actual target gas concentration; The corrected target gas concentration Ccorrected is output as the true concentration estimate at the current time point.

8. The method for evaluating the filtering effect of semiconductor manufacturing waste gas according to claim 7, characterized in that: The purification efficiency calculation formula of the semiconductor exhaust gas filtration system is: ; Where η is the purification efficiency, is the target gas concentration before the exhaust gas enters the filtration system, is the target gas concentration after the exhaust gas is treated; For each set of sampling data, the purification efficiency value is calculated based on the corrected inlet and outlet concentrations; The purification efficiency results are displayed visually. If the efficiency is lower than the set threshold, an early warning prompt is triggered to assist maintenance personnel in making judgments and optimization adjustments.

9. A semiconductor manufacturing waste gas filtration effect evaluation system, used to implement a semiconductor manufacturing waste gas filtration effect evaluation method according to any one of claims 1 to 8, characterized in that: It includes data acquisition module, feature extraction module, error prediction module, correction module and purification efficiency evaluation module; Data acquisition module: collects the response data obtained by multiple gas sensors of different types for synchronous detection of the same exhaust gas to be tested; Feature extraction module: extracts gas interference features from the response data of different sensors. The interference features include abnormal spectrum morphology features and time-varying gradient features between sensors. Error prediction module: Based on historical experimental data and known interference gas impact mechanisms, a machine learning algorithm is used to train a cross-interference prediction model. The real-time collected gas interference features are input into the cross-interference prediction model to predict the severity of errors caused by cross-interference in the current monitoring data. Correction module: According to the severity of the error, different gas concentration correction strategies are selected to perform differential correction on the original monitoring data of the target gas to obtain the target gas concentration value after cross-interference correction; Purification efficiency evaluation module: Based on the corrected target gas concentration value, the purification efficiency of the semiconductor exhaust gas filtration system is evaluated.

Images