A fault detection and classification system for semiconductors

Abstract

This invention discloses a fault detection and classification system for semiconductors, relating to the field of fault detection technology. It collects raw data of semiconductor process parameters in real time using sensors; divides the collected raw data into multiple sub-step data segments according to a predetermined time period; and extracts features from each data segment based on its distribution characteristics and distribution characteristic index. The extracted feature data is then input into a preset quality inspection model to inspect semiconductor products, outputting the fault type of the wafer. This system can judge complex data distributions and dynamically adjust the feature extraction algorithm to adapt to different data distribution characteristics, effectively extracting potentially important features from the data and ensuring the accuracy of fault detection. Especially in complex process environments, it reduces the probability of misjudgment or missed judgment, minimizing the impact on the quality and efficiency of semiconductor production.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and more specifically to a fault detection and classification system for semiconductors. Background Technology

[0002] In semiconductor manufacturing, the operating status of production equipment is typically monitored in real time using numerous sensors. The raw data collected includes various process parameters such as temperature, pressure, flow rate, and power. This data is usually time-series data, exhibiting strong dynamic variability and complexity. Traditional fault detection methods often rely on feature extraction from this time-series data to identify abnormal patterns or trends. Traditional methods employ fixed feature extraction algorithms, such as calculating statistical features like mean, standard deviation, maximum, and minimum values, but this approach has significant limitations.

[0003] In practical applications, semiconductor process data often exhibit different distribution characteristics. For example, some data may show periodic changes, such as sensor data in etching and ion implantation processes, which typically show periodic fluctuations; while other data may show characteristics of being concentrated in a certain range, such as certain machine parameters remaining stable for a long time with small fluctuations. Therefore, traditional fixed algorithms are difficult to effectively adapt to different data distribution characteristics, resulting in low accuracy and efficiency in fault detection.

[0004] Existing fault detection methods not only ignore the complex distribution of data, but also fail to dynamically adjust feature extraction algorithms to adapt to different data distribution characteristics, thus failing to effectively extract potentially important features from the data. This directly affects the accuracy of fault detection, especially in complex process environments, easily leading to false positives or false negatives, thereby impacting the quality and efficiency of semiconductor production. Summary of the Invention

[0005] The purpose of this invention is to solve the problems mentioned above and provide a fault detection and classification system for semiconductors.

[0006] This invention proposes a fault detection and classification system for semiconductors, the system comprising:

[0007] Data acquisition module: Collects raw data of semiconductor process parameters in real time through sensors; and divides the collected raw data into data segments of multiple sub-steps according to a predetermined time period;

[0008] Distribution Feature Module: Calculates the distribution feature index of each data segment based on the distribution feature data of each segment after segmentation;

[0009] Feature extraction module: Based on the distribution characteristic index, select the corresponding feature extraction algorithm to extract features from the data segment;

[0010] Fault Detection and Classification Module: Extracts feature data from the data segment and inputs it into the preset quality inspection model to perform quality inspection on semiconductor products. It outputs the fault detection results and fault types of the wafer, determines whether there are faults that cause the wafer to be unqualified, and classifies and identifies the fault types.

[0011] Optionally, the raw data includes temperature, pressure, gas flow rate, and power; the data segment of each sub-step represents a different stage in the semiconductor manufacturing process, and the data segment of each sub-step forms a time series.

[0012] Optionally, the step of calculating the distribution characteristic index of each data segment based on the distribution characteristic data of each segment after segmentation is as follows:

[0013] The distribution feature data includes a scale distribution index and an information geometric metric index. The scale distribution index and the information geometric metric index are normalized, and the normalized scale distribution index is subtracted from the normalized information geometric metric index to obtain the distribution feature index.

[0014] Optionally, the calculation steps for the scale distribution index are as follows:

[0015] The collected raw data is subjected to first-order difference processing to obtain the difference sequence between every two adjacent data points;

[0016] The first-order difference sequence is decomposed into multiple scales using wavelet transform, and the data energy at each scale is calculated based on the wavelet transform results.

[0017] Calculate the autocorrelation coefficient of the data at each scale, divide the data segment corresponding to each scale into several windows, calculate the difference between all two adjacent data points within the window for each window, and sum the product of all differences as the similarity of each window; average the similarity results of all windows to obtain the autocorrelation coefficient of the corresponding scale.

[0018] The energy and self-similarity index at each scale are normalized; the normalized energy and self-similarity index at each scale are multiplied together, and the sum of the results of multiplying the normalized energy and self-similarity index at all scales is taken as the scale distribution index.

[0019] Optionally, the calculation steps for the information geometric metric index are as follows:

[0020] The original data of each data segment is standardized, and the standardized data of each data segment is mapped to a high-dimensional space through nonlinear mapping, and the covariance matrix is ​​calculated in the high-dimensional space.

[0021] Based on the covariance matrix, the Riemannian measure of the data in high-dimensional space is calculated; and based on the Riemannian measure, the local curvature of each data point is calculated.

[0022] The local curvature and the Riemannian metric are added together to obtain the local geometric complexity of each data point;

[0023] The mean of the local geometric complexity of all data points is taken as the global geometric complexity of the corresponding data segment. The global geometric complexity of each data segment is then normalized, and the result of the normalization is taken as the information geometric metric index.

[0024] Optionally, the steps for selecting the corresponding feature extraction algorithm to extract features from the data segment based on the distribution characteristic index are as follows:

[0025] Compare the distribution characteristic index with the preset distribution characteristic index threshold. If the distribution characteristic index is not less than the preset distribution characteristic index threshold, then the data segment has periodic distribution characteristics.

[0026] If the distribution characteristic index is less than the preset distribution characteristic index threshold, then the data segment has a distribution characteristic that is concentrated in a certain interval;

[0027] Select the appropriate feature extraction algorithm to extract features from the data segment based on its different distribution characteristics.

[0028] Optionally, the steps for selecting the appropriate feature extraction algorithm to extract features from the data segment based on its different distribution characteristics are as follows:

[0029] If the data segment has a periodic distribution characteristic, then frequency domain analysis algorithm and statistical probability distribution algorithm are used to extract features;

[0030] If the data segment has a distribution characteristic concentrated in a certain interval, then the extreme value analysis algorithm is used to extract features.

[0031] The beneficial effects of this invention are:

[0032] This invention proposes a fault detection and classification system for semiconductors. It collects raw data of semiconductor process parameters in real time using sensors; divides the collected raw data into multiple sub-step data segments according to a predetermined time period; calculates the distribution characteristic index of each data segment based on its distribution characteristics; and selects a corresponding feature extraction algorithm to extract features from the data segments based on the distribution characteristic index. The extracted feature data is then input into a preset quality inspection model to inspect semiconductor products, outputting the wafer fault detection results and fault types. This system determines whether there are faults causing wafer defects and classifies the fault types. In this way, it can judge complex data distributions and dynamically adjust the feature extraction algorithm to adapt to different data distribution characteristics, thereby effectively extracting potentially important features from the data and ensuring the accuracy of fault detection. Especially in complex process environments, it reduces the probability of false positives or false negatives, minimizing the impact on the quality and efficiency of semiconductor production. Attached Figure Description

[0033] The invention will now be further described with reference to the accompanying drawings.

[0034] Figure 1 This is a framework diagram of a fault detection and classification system for semiconductors. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0037] This invention provides a fault detection and classification system for semiconductors. See also... Figure 1 , Figure 1 A flowchart illustrating a fault detection and classification system for semiconductors, provided as an embodiment of the present invention. The system includes:

[0038] Data acquisition module: Real-time acquisition of raw data of semiconductor process parameters through sensors; and the acquired raw data is divided into multiple data segments of sub-steps according to a predetermined time period. Each data segment of sub-steps represents a different stage in the semiconductor manufacturing process, and each data segment of sub-steps forms a time series; the raw data includes, but is not limited to, process parameters such as temperature, pressure, gas flow rate, and power.

[0039] Distribution Feature Module: Based on the distribution feature data of each segmented data segment, calculate the distribution feature index of each data segment; the distribution feature index is used to reflect whether the data segment has periodic distribution characteristics or is concentrated in a certain interval.

[0040] Feature extraction module: Based on the distribution characteristic index, select the corresponding feature extraction algorithm to extract features from the data segment;

[0041] Fault Detection and Classification Module: Extracts feature data from the data segment and inputs it into the preset quality inspection model to perform quality inspection on semiconductor products. It outputs the fault detection results and fault types of the wafer, determines whether there are faults that cause the wafer to be unqualified, and classifies and identifies the fault types.

[0042] Based on the fault detection and classification system for semiconductors provided in this embodiment of the invention, the system can judge the complex distribution of data and dynamically adjust the feature extraction algorithm to adapt to the distribution characteristics of different data, thereby effectively extracting potential important features in the data and ensuring the accuracy of fault detection. Especially in complex process environments, it reduces the probability of misjudgment or omission, and further reduces the impact on the quality and efficiency of semiconductor production.

[0043] In one embodiment, the data acquisition module: acquires raw data of semiconductor process parameters in real time through sensors; and divides the acquired raw data into multiple data segments of sub-steps according to a predetermined time period. Each data segment of a sub-step represents a different stage in the semiconductor manufacturing process, and each data segment of a sub-step forms a time series. The raw data includes, but is not limited to, process parameters such as temperature, pressure, gas flow rate, and power.

[0044] It should be noted that during the data acquisition step, the system collects raw data of semiconductor process parameters in real time through sensors. These parameters reflect the status of each stage of the production process and provide necessary information for subsequent fault detection and classification. The process data collected by the sensors includes, but is not limited to, parameters such as temperature, pressure, gas flow rate, and power. These parameters reflect the dynamic changes of different equipment and process steps in the manufacturing process. For example, temperature may be used to monitor the heating process in semiconductor production, pressure to track gas flow in the etching process, gas flow rate to reflect changes in gas injection volume, and power to reflect equipment load and energy efficiency.

[0045] The collected raw data is segmented according to predetermined time periods, with each time period corresponding to a specific process step in the production process. This segmentation aims to break down the data into sub-step data segments related to different process stages. These sub-step data segments represent the operational status of different stages in semiconductor manufacturing, such as the operating status of a specific process (e.g., etching, doping) or equipment (e.g., reactor, lithography machine). Each sub-step data segment forms a time series, meaning that the parameter data within each time period are arranged chronologically to analyze their trends over time.

[0046] For example, in a channel etching process, sensors may record front-end pressure, temperature, and gas flow data at different time points. This data is segmented according to the various stages of the etching process (such as plasma activation, etching reaction, and gas conditioning), with each stage's data segment forming an independent time series representing the dynamic changes in process parameters at that stage. These data series not only help identify the operational status of each process stage but also provide essential input for subsequent data analysis and fault detection.

[0047] In one embodiment, the distribution feature module calculates the distribution feature index of each data segment based on the distribution features of each segment after segmentation; the distribution feature index is used to reflect whether the data segment has periodic distribution features or is concentrated in a certain interval.

[0048] In one implementation, the step of calculating the distribution characteristic index of each data segment based on the distribution characteristic data of each segment after segmentation is as follows:

[0049] The distribution characteristic data includes the scale distribution index and the information geometric metric index. The scale distribution index and the information geometric metric index are normalized, and the normalized scale distribution index is subtracted from the normalized information geometric metric index to obtain the distribution characteristic index.

[0050] The normalization methods mentioned above include Min-Max normalization, Z-Score standardization, etc., which will not be elaborated here.

[0051] In one embodiment, the steps for calculating the scale distribution index are as follows:

[0052] The original data of each data segment is subjected to first-order differencing to obtain the difference sequence between every two adjacent data points. This operation helps to reveal the local fluctuation characteristics of the data segment and provides a more stable data structure for subsequent scaling.

[0053] Wavelet transform is used to decompose the obtained first-order difference sequence into multiple scales, and the data energy at each scale is calculated based on the wavelet transform results. This step maps the difference sequence to multiple scales to reveal the variation characteristics of the data at different scales. Wavelet transform converts the first-order difference sequence into frequency domain features at different scales, reflecting the fluctuation behavior of the data at each scale and providing a basis for subsequent energy calculation. The decomposition result at each scale represents the variation information of the data at different scales. Energy reflects the fluctuation intensity of the data at that scale; a higher energy value indicates greater data variation and stronger fluctuation at that scale. By calculating the energy at each scale, the contribution of each scale to the overall data fluctuation can be quantified.

[0054] To calculate the autocorrelation coefficient of the data at each scale, the data segment corresponding to each scale is divided into several windows. For each window, the difference between all two adjacent data points within the window is calculated, and the product of all differences is accumulated to obtain the similarity of each window. The similarity results of all windows are averaged to obtain the autocorrelation coefficient of the corresponding scale. If the coefficient is high, it indicates that the data has strong self-similarity at that scale; if the coefficient is low, it indicates that the data varies greatly at that scale and lacks regularity.

[0055] The energy and self-similarity index at each scale are normalized. The normalized energy and self-similarity index at each scale are then multiplied, and the sum of these multiplications across all scales is used as the scale distribution index. This index comprehensively considers the intensity and regularity of data fluctuations at different scales, reflecting the periodicity of the data. A higher scale distribution index indicates stronger periodicity in the data.

[0056] It's important to note that the scale distribution index is a comprehensive quantitative indicator used to measure the fluctuation characteristics and self-similarity of data segments at different scales, aiming to reveal the periodic or concentrated characteristics of the data. Multi-scale analysis allows us to explore the regular changes in data across various scales. Specifically, a larger scale distribution index indicates stronger self-similarity and more regular fluctuation patterns across multiple scales, typically suggesting a periodic distribution. Conversely, a smaller scale distribution index indicates a lack of regularity and consistency across scales, usually suggesting that the data fluctuates within a specific range and exhibits concentrated distribution characteristics. When the scale distribution index is large, it means that the fluctuations of data at different scales can better reflect similar change patterns; this repetitiveness often appears in periodic data. For example, in periodic phenomena, data fluctuations maintain a certain frequency or shape over time; therefore, the energy distribution and self-similarity indicators of the data will show high consistency across multiple scales. For example, suppose we are observing a periodic oscillation process (such as a mechanical vibration signal). The fluctuation pattern of this signal reflects similar periodic characteristics at different time scales. Therefore, the energy and self-similarity at each scale will show high consistency, resulting in a large scale distribution index. Conversely, when data is concentrated in a certain interval, the fluctuations are more localized and the changes are more random, lacking periodicity. For example, in a certain production process, the operating state of the equipment is relatively stable, and data such as temperature and pressure will fluctuate within a small range, lacking obvious regularity and periodicity. In this case, the changes in data at different scales will appear more irregular, with lower energy distribution and self-similarity indices, resulting in a smaller scale distribution index, indicating that the data exhibits a concentrated distribution. Therefore, a larger scale distribution index indicates that the data exhibits strong periodicity and consistency across multiple scales, conforming to the characteristics of a periodic distribution; while a smaller scale distribution index indicates that the data is more likely to exhibit localized fluctuations concentrated in a certain interval, conforming to the characteristics of a concentrated distribution.

[0057] It should be noted that the above method for calculating the scale distribution index combines multiple complex steps, including first-order differencing, wavelet transform, autocorrelation index, and normalization, effectively capturing the changing patterns of data from multiple dimensions. Compared with other calculation methods, the biggest advantage of this method is that it comprehensively considers local fluctuations, multi-scale features, and data self-similarity, thus revealing the periodic or concentrated distribution characteristics of the data more accurately. First, first-order differencing can eliminate long-term trends or noise in the data, focusing only on local fluctuations, which is very effective in capturing periodic features. Then, multi-scale decomposition through wavelet transform can extract local features of the data at different scales, which is particularly important for identifying the fluctuation behavior of data at multiple scales, especially when data changes are not limited to a single scale. Next, by calculating the autocorrelation index, the self-similarity of the data at different scales can be further quantified, helping to determine whether the data has periodicity. Autocorrelation is an important indicator for judging whether data has recurring patterns or regularity; this step can effectively remove fluctuations that lack periodicity or stability and exclude concentrated data. The normalization step helps eliminate scale bias, ensuring the comparability of features at different scales and preventing the impact of excessively large data volumes at certain scales. In this way, data energy and self-similarity at all scales can be calculated under the same dimensions, resulting in a scale distribution index that comprehensively reflects the global fluctuation characteristics of the data. In general, the advantage of this calculation method is that it comprehensively captures the periodic and central characteristics of the data, avoiding the shortcomings of relying solely on a single scale or simple statistical methods. Through multi-level, multi-scale analysis, the resulting scale distribution index has higher robustness and accuracy, more accurately reflecting the true distribution characteristics of the data, and is particularly suitable for complex data with multiple fluctuation patterns.

[0058] In one embodiment, the calculation steps for the information geometric metric index are as follows:

[0059] The original data of each data segment is standardized, and the standardized data of each data segment is mapped to a high-dimensional space through a nonlinear mapping (such as Gaussian kernel function), and the covariance matrix is ​​calculated in the high-dimensional space.

[0060] Standardization ensures that the original data for each segment has a mean of 0 and a variance of 1. The standardized dataset removes the influence of different dimensions or units, making subsequent geometric analysis more consistent. Nonlinear mapping helps reveal more complex relationships and structures in the data, allowing the data to exhibit its geometric characteristics more precisely in high-dimensional space. The covariance matrix reflects the degree of variation in the data across various dimensions. Specifically, each element of the covariance matrix represents the linear relationship between two dimensions, and the matrix's properties reveal the global geometric characteristics of the data.

[0061] Based on the covariance matrix, the Riemannian metric of the data in high-dimensional space is calculated; and based on the Riemannian metric, the local curvature of each data point is calculated. The Riemannian metric reflects the geometric distance of the data in this space. The Riemannian metric measures the similarity or difference between data points by calculating the geometric distance between them. Curvature reflects the rate of change of a data point in high-dimensional space, representing the complexity of the data near that point. By calculating the curvature of each data point, the intensity of local changes in the data can be obtained. A large local curvature indicates a complex data structure around that point and significant local changes; a small local curvature indicates that the data changes gradually near that point.

[0062] The local geometric complexity of each data point is obtained by adding the local curvature and the Riemannian metric; this value measures the geometric complexity of the region where the data point is located. Local geometric complexity reflects the geometric complexity of the data surrounding that point by analyzing the rate of change of local curvature. Higher local geometric complexity indicates that the data in that region is complex, exhibiting irregularity or local concentration; while lower geometric complexity indicates that the data in that region is relatively simple and regular.

[0063] The mean of the local geometric complexity of all data points is taken as the global geometric complexity of the corresponding data segment. The global geometric complexity of each data segment is then normalized, and the result of the normalization is taken as the information geometric metric index.

[0064] Global geometric complexity reflects the geometric properties of the entire dataset, comprehensively considering the complexity of all local regions. A higher global geometric complexity indicates a more complex geometric structure and potentially more irregularities; conversely, a lower global geometric complexity indicates a simpler structure and potentially stronger periodicity. The information geometric metric index comprehensively evaluates the geometric structural characteristics of the data, reflecting its complexity and regularity. A smaller information geometric metric index indicates a simpler and more regular structure in geometric space, typically corresponding to a periodic distribution; conversely, a larger index indicates a more complex and irregular structure, typically corresponding to a concentrated distribution.

[0065] It's important to note that the information geometric metric is a crucial indicator for measuring the geometric complexity and regularity of data in high-dimensional space. It quantifies the distribution characteristics of data by combining multi-dimensional geometric information such as the covariance matrix, Riemannian metric, local curvature, and geometric complexity. Specifically, the information geometric metric reflects the distribution structure and variation patterns of data in high-dimensional space. A smaller information geometric metric indicates a simpler and more regular geometric structure, typically corresponding to a periodic distribution. Periodic data usually exhibits a relatively regular structure and low curvature in high-dimensional space, meaning that the data changes with strong regularity and relatively stable overall fluctuations. Conversely, a larger information geometric metric indicates a more complex geometric structure and greater local variation in high-dimensional space, usually indicating a concentrated distribution. Concentrated data often exhibits high local curvature, with significant fluctuations within a certain range, lacking obvious regularity or periodicity. For example, under stable operating conditions, data such as temperature or pressure may fluctuate within a small range, lacking periodicity. Such data would exhibit a large information geometric metric, reflecting high local variation and significant geometric complexity. In this way, the information geometric metric index provides an effective tool for determining whether data has periodic characteristics, thereby providing a basis for subsequent fault detection and classification.

[0066] It's important to note that calculating the information geometric metric index using the above method offers advantages over other methods. It comprehensively captures the complexity and regularity of data from the perspective of high-dimensional space and geometric structure, particularly effectively identifying whether the data exhibits periodic or concentrated distribution characteristics. First, mapping data to a high-dimensional space through nonlinear mapping overcomes the limitations of traditional linear analysis methods, allowing the nonlinear characteristics of the data to be effectively expressed in high-dimensional space. This is crucial for revealing the intrinsic relationships within complex data. Second, calculating the covariance matrix and Riemannian metric allows us to focus not only on simple distances or correlations but also on understanding the geometric structure and topological features of the data in high-dimensional space, thereby quantifying the data's distribution characteristics. This metric reveals the geometric distances between data points, far exceeding simple statistical analysis methods considering the complex patterns of data variation. Furthermore, calculating local curvature quantifies the nonlinear changes and local fluctuations of the data, reflecting the complexity of the data in different regions. Finally, by calculating local geometric complexity, we can comprehensively consider the geometric characteristics of each data point to obtain an overall distribution characteristic index. The normalized geometric metric index, ranging from 0 to 1, provides a unified and easily comparable indicator for the data. Overall, this method can more accurately and comprehensively measure the geometric complexity of data, especially when dealing with data exhibiting complex patterns and multi-dimensional variations. It offers more accurate and reliable distributional characteristic analysis, proving far more effective than traditional simple statistical methods.

[0067] In one embodiment, the feature extraction module: selects the corresponding feature extraction algorithm to extract features from the data segment based on the distribution feature index;

[0068] Specifically, the steps for selecting the appropriate feature extraction algorithm to extract features from a data segment based on the distribution characteristic index are as follows:

[0069] Compare the distribution characteristic index with the preset distribution characteristic index threshold. If the distribution characteristic index is not less than the preset distribution characteristic index threshold, then the data segment has periodic distribution characteristics.

[0070] If the distribution characteristic index is less than the preset distribution characteristic index threshold, then the data segment has a distribution characteristic that is concentrated in a certain interval;

[0071] Select the appropriate feature extraction algorithm to extract features from the data segment based on its different distribution characteristics.

[0072] In one implementation, the steps for selecting the appropriate feature extraction algorithm to extract features from the data segment based on its different distribution characteristics are as follows:

[0073] If the data segment has a periodic distribution characteristic, then frequency domain analysis algorithm and statistical probability distribution algorithm are used to extract features;

[0074] If the data segment has a distribution characteristic concentrated in a certain interval, then use extreme value analysis algorithms (such as maximum value, minimum value) to extract features.

[0075] It should be noted that in the step of selecting a feature extraction algorithm, the first step is to compare the calculated distribution characteristic index with a preset distribution characteristic index threshold to determine the distribution type of the data segment. If the distribution characteristic index is not less than the preset threshold, it indicates that the data segment has periodic distribution characteristics, and a feature extraction algorithm suitable for periodic data is selected. If the distribution characteristic index is less than the preset threshold, it indicates that the fluctuation of the data segment is relatively localized, belonging to the distribution characteristics concentrated in a certain interval, and a feature extraction algorithm specifically for concentrated data is selected.

[0076] Specifically, for data segments with periodic distribution characteristics, frequency domain analysis algorithms can be used to extract the frequency components of the data, helping to identify periodic patterns. For example, using Fourier transform to convert the data from the time domain to the frequency domain yields the intensity distribution of different frequency components, revealing the periodic variation patterns of the data. Furthermore, statistical probability distribution algorithms can be combined, such as calculating the mean, variance, and standard deviation of the data, to further confirm the regularity and periodicity of the data fluctuations. For example, in periodic data, we might find that the data fluctuations conform to a certain normal distribution or other known distribution types, thus helping us better understand the regularity of the data. For data segments with distribution characteristics concentrated in a certain interval, extreme value analysis algorithms (such as maximum and minimum values) are used to extract the extreme value characteristics of the data. Since the fluctuations of concentrated data are relatively localized, the range of change of data points is usually limited; maximum and minimum values ​​can effectively reflect the boundaries of data change. For example, in a temperature-controlled production process, if the data is consistently concentrated within a small range, using maximum and minimum values ​​can reveal the upper and lower limits of the data, thereby helping to determine whether the equipment is operating within a stable operating range. In this way, the selection of feature extraction algorithms is closely related to the distribution characteristics of the data, ensuring the accuracy of feature extraction and the full representation of the data. This algorithm selection based on distribution characteristics can flexibly adapt to the characteristics of different types of data, avoiding the limitations of applying a single algorithm in traditional methods, and improving the accuracy and efficiency of fault detection and classification.

[0077] In one embodiment, the fault detection and classification module extracts feature data from the data segment and inputs it into a preset quality inspection model to perform quality inspection on semiconductor products, outputs the fault detection results and fault types of the wafer, determines whether there are faults that cause the wafer to be unqualified, and classifies and identifies the fault types.

[0078] It's important to note that in the fault detection and classification step, the feature data extracted from the data segment is first input into a pre-defined quality inspection model for analysis. This quality inspection model is typically a machine learning model trained using historical data and known fault patterns, such as Support Vector Machine (SVM), Random Forest (RF), or neural networks. These models can determine the wafer's quality status based on the input feature data. By inputting the feature data, the quality inspection model assesses whether the wafer meets predetermined quality standards and provides fault detection results. Specifically, the input data for the quality inspection model usually includes features extracted from collected process parameters (such as temperature, pressure, gas flow rate, and power). These features reflect potential quality problems in the wafer during production. For example, if temperature or pressure fluctuations exceed normal ranges in a certain production stage, the feature data may contain this fluctuation information, serving as input to the quality inspection model. Based on this input data, the quality inspection model outputs a wafer quality assessment result and determines whether a fault exists. The output results may include: Fault detection result: Is the wafer qualified? If the wafer meets production standards, it is marked as "qualified"; if an anomaly occurs, it is marked as "unqualified". Fault Types: If a wafer is defective, the model will further analyze and identify the fault type based on the input data. Fault types may include, but are not limited to: Dimensional issues: such as the wafer's thickness or diameter exceeding specified ranges. Material defects: such as scratches, cracks, or other material quality problems on the wafer surface. Process anomalies: such as production defects caused by abnormal temperature or pressure fluctuations. For example, suppose that during the photolithography process, an irregular crack appears on the surface of a batch of wafers due to excessive temperature. In this case, the temperature fluctuation and crack information in the feature data will be input into the quality inspection model. The model will identify the abnormal fluctuation and output the fault type as "material defect caused by temperature". If the fault type does not conform to the preset quality standard, the wafer will be judged as "unqualified". Ultimately, by analyzing the output fault detection results and fault types, the quality inspection model helps production line workers quickly identify and locate potential problems, ensuring that only wafers that meet quality standards enter the next stage of processing.

[0079] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A fault detection and classification system for a semiconductor, characterized by, The system comprises: a data acquisition module: collecting raw data of semiconductor process parameters in real time through sensors; and cutting the collected raw data into data segments of multiple sub-steps according to a predetermined time period; a distribution feature module: calculating distribution feature indexes of the data segments according to distribution feature data of the cut data segments; the specific process is as follows: performing first-order difference processing on the collected raw data to obtain a difference value sequence between every two adjacent data points; performing multi-scale decomposition on the obtained first-order difference sequence using wavelet transform, and calculating data energy at each scale according to the result of wavelet transform; calculating data autocorrelation coefficients at each scale, dividing the data segment corresponding to each scale into a plurality of windows, for each window, calculating the difference value between all two adjacent data points in the window, and adding the product of all the difference values as the similarity of each window; and averaging the similarity results of all the windows to obtain the autocorrelation coefficient of the corresponding scale; normalizing the energy and self-similarity index at each scale; multiplying the normalized energy and self-similarity index at each scale, and taking the sum of the multiplication results of all the normalized energy and self-similarity indexes at all the scales as the scale distribution index; a feature extraction module: selecting a corresponding feature extraction algorithm to extract features from the data segment according to the distribution feature index; the specific process is as follows: if the data segment is of periodic distribution feature, using a frequency domain analysis algorithm and a statistical probability distribution algorithm to extract features; if the data segment is of distribution feature concentrated in a certain interval, using an extreme value analysis algorithm to extract features; a fault detection and classification module: inputting the feature data extracted from the data segment into a preset quality inspection model to perform quality inspection on the semiconductor product, outputting the fault detection result and the fault type of the wafer, judging whether there is a fault causing the wafer to be unqualified, and classifying and identifying the fault type.

2. A fault detection and classification system for a semiconductor as claimed in claim 1, wherein, The raw data includes temperature, pressure, gas flow and power; the data segment of each sub-step represents different stages in the semiconductor manufacturing process, and each data segment of each sub-step forms a time sequence.

3. A fault detection and classification system for a semiconductor as claimed in claim 1, wherein, The step of calculating the distribution feature index of each data segment according to the distribution feature data of the cut data segments is as follows: The distribution feature data includes a scale distribution index and an information geometric metric index, the scale distribution index and the information geometric metric index are normalized, and the normalized scale distribution index is subtracted from the normalized information geometric metric index to obtain the distribution feature index.

4. A fault detection and classification system for a semiconductor as claimed in claim 3, wherein, The calculation steps of the information geometric metric index are as follows: standardizing the raw data of each data segment, mapping the standardized data of each data segment to a high-dimensional space through nonlinear mapping, and calculating a covariance matrix in the high-dimensional space; based on the covariance matrix, calculating the Riemannian metric of the data in the high-dimensional space; and based on the Riemannian metric, calculating the local curvature of each data point; adding the local curvature and the Riemannian metric to obtain the local geometric complexity of each data point; The mean of the local geometric complexity of all data points is taken as the global geometric complexity of the corresponding data segment, and the global geometric complexity of each data segment is normalized, and the normalized result is taken as the information geometric metric index.

5. The fault detection and classification system for a semiconductor as claimed in claim 1, wherein, According to the distribution characteristic index, a corresponding feature extraction algorithm is selected to extract features from the data segment. The distribution characteristic index is compared with the preset distribution characteristic index threshold value, and if the distribution characteristic index is not less than the preset distribution characteristic index threshold value, the data segment has a periodic distribution characteristic; If the distribution characteristic index is less than the preset distribution characteristic index threshold value, the data segment has a distribution characteristic of being concentrated in a certain interval; According to the different distribution characteristics of the data segment, a corresponding feature extraction algorithm is selected to extract features from the data segment.

Images