A Method and System for Cavity Health Analysis Based on Big Data Statistics

Abstract

This invention discloses a cavity health analysis method and system based on big data statistics, primarily applied to cavity equipment monitoring in fields such as semiconductor manufacturing. The method first collects historical and real-time data of the cavity, including yield rate characteristics, component maintenance dimensions, and FDC data dimensions. Then, it calculates the health score for each dimension: the yield rate health score (H1) is obtained by averaging the normalized numbers of four types of defects; the component maintenance health score (H2) is obtained by multiplying the scores of each component after evaluating them according to preset rules; and the FDC health score (H3) is calculated based on the product of the complements of the alarm rate of the key parameter. Finally, a comprehensive cavity health score (S) is generated by weighting and fusing the health scores of the three dimensions based on preset weights. When the S value is below a threshold, a maintenance warning and shutdown command are automatically triggered, achieving multi-dimensional data fusion analysis and significantly improving the accuracy of cavity status assessment and maintenance efficiency.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing, and in particular relates to a method and system for cavity health analysis based on big data statistics. Background Technology

[0002] In semiconductor manufacturing, the health status of the chamber is crucial to wafer processing yield. Traditional chamber health analysis methods heavily rely on engineers' expert experience and basic mathematical statistics. Existing technologies typically employ fragmented assessments: either monitoring only the quantity changes of yield defects (such as True defects, Repeater defects, etc.) and issuing warnings through simple thresholds (such as triggering maintenance when defects exceed limits), or generating discrete recommendations solely based on component maintenance records (such as maintenance frequency and number of failures), or independently analyzing alarm rates for FDC (Fault Detection and Classification) parameters (such as pressure and temperature fluctuations). This fragmented approach leads to one-sided health assessments, low accuracy, and poor robustness. Especially for new engineers, the experience barrier significantly reduces work efficiency—for example, independent dimensional analysis cannot capture the synergistic effects of yield decline and component wear or FDC alarms, often leading to misjudgments or delayed responses, resulting in unplanned downtime, yield fluctuations, and increased operating costs. The main technical challenge lies in how to establish a cross-dimensional, dynamically integrated cavity health assessment model that replaces reliance on human experience with data-driven approaches, thereby achieving accurate quantification and real-time early warning of systemic health risks. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a cavity health analysis method and system based on big data statistics to address the shortcomings of the prior art. It establishes a cross-dimensional dynamic fusion cavity health assessment model, replaces reliance on manual experience with data-driven approaches, and achieves accurate quantification and real-time early warning of systemic health risks.

[0004] The first aspect of this invention discloses a method for analyzing the health status of bodily cavities based on big data statistics, characterized by comprising the following steps:

[0005] Historical and real-time data of the cavity are collected, including yield dimension characteristics, component maintenance dimension, and FDC data dimension.

[0006] Based on the aforementioned yield dimension features, calculate the yield dimension health score. The yield dimension features include the number of True defects, Repeater defects, Big defects, and Spatial signature defects, calculated as follows: for each type of defect... Perform normalization to obtain the historical maximum value. and historical minimum Then the health value of a single defect is Then, the average of the four defect health values ​​is obtained. ;

[0007] Calculate the component maintenance health level based on the aforementioned component maintenance dimensions. The component maintenance dimensions include maintenance frequency, most recent maintenance interval, number of failures, and average maintenance cycle. The calculation method is as follows: Each component's health score is determined according to preset rules. ,in These represent the health scores of different components;

[0008] Calculate the health status of the FDC data dimension based on the aforementioned FDC data dimension. The FDC data dimensions include the standard deviation of pressure fluctuations, the alarm rate for exceeding temperature fluctuation limits, and the alarm rate for large vacuum fluctuations. The calculation method is as follows ;

[0009] Based on weight ,in Calculate the overall health score of the cavity. .

[0010] In the above method, the Spatial signature defect includes scratch defects or annular defects, whose health value is calculated independently of other defects, and the normalized weight of scratch defects is 1.2 times that of other defects.

[0011] The above method uses Z-score normalization for normalization, and the health value for a single defect is calculated as follows: ,in This is the historical average. The standard deviation is the historical value.

[0012] The above method, the weight Based on real-time alarm rate dynamic adjustment, the calculation method is as follows: ,in For the first Current alarm rate for the dimension.

[0013] The above method also includes the step of: when the overall health score... When the value drops below 0.5, a maintenance warning will be automatically triggered and a shutdown command will be generated.

[0014] In the above method, the alarm threshold of the FDC data dimension is user-defined, including pressure fluctuation threshold (±0.3psi) or temperature fluctuation range (±2℃).

[0015] The above method is applied to multiple cavities and calculates a factory-level average health score. ,in For the first Health score for each cavity.

[0016] The above method includes a data acquisition step that involves acquiring data in real time from a sensor or FDC system via an API interface and filling missing values ​​with linear interpolation.

[0017] A second aspect of this invention discloses a cavity health analysis system, comprising:

[0018] The data acquisition module is used to acquire the yield dimension characteristics, component maintenance dimension, and FDC data dimension of the cavity in real time;

[0019] The health score calculation module includes:

[0020] The yield analysis unit is configured to calculate the number of four types of defects: True defect, Repeater defect, Big defect, and Spatial signature defect. ; Number of defects The historical maximum value, Number of defects The historical minimum value;

[0021] The component analysis unit is configured to output based on preset rules. ;in These represent the health scores of different components;

[0022] The FDC analysis unit is configured to calculate based on the alarm rate. ;in, These are the alarm rates for pressure fluctuation standard deviation, temperature fluctuation exceeding limits, and large vacuum fluctuation, respectively.

[0023] The comprehensive evaluation module is configured to be weight-based. Calculate the overall health score of the cavity ;

[0024] The early warning execution module is configured to... Maintenance commands are triggered at certain times.

[0025] The third aspect of this invention discloses a cavity health analysis software system, running on a server cluster, comprising:

[0026] The data interface unit is used to receive yield, component maintenance, and FDC data from the sensor network;

[0027] Process the engine and perform the following operations:

[0028] For the yield data, calculate and output the normalized health value. The yield dimension features include the number of True defects, Repeater defects, Big defects, and Spatial signature defects, calculated as follows: for each type of defect... Perform normalization to obtain the historical maximum value. and historical minimum Then the health value of a single defect is Then, the average of the four defect health values ​​is obtained. ;

[0029] Maintenance data for parts is generated based on a preset rule table. The component maintenance dimensions include maintenance frequency, most recent maintenance interval, number of malfunctions, and average maintenance cycle. The calculation method is as follows: Each component's health score is determined according to preset rules. ,in These represent the health scores of different components;

[0030] Calculate the product output for FDC alarm rate data. The FDC data dimensions include the standard deviation of pressure fluctuations, the alarm rate for exceeding temperature fluctuation limits, and the alarm rate for large vacuum fluctuations. The calculation method is as follows ;

[0031] Based on dynamic weight formula Calculate the overall score of the cavity ;

[0032] Control command unit, when When the value falls below the threshold, a shutdown signal is sent to the maintenance terminal.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] 1. By employing big data statistical methods (such as historical data normalization and complementary calculation of alarm rates), stable patterns can be extracted from massive amounts of data. For example, in the yield dimension, historical baseline (max / min) calibration is performed for four types of defects, including True defects, avoiding misjudgments caused by a single threshold; regularized scoring in the component maintenance dimension (based on frequency, number of failures, etc.) ensures the objectivity of the scoring; and complementary multiplication of alarm rates in the FDC dimension (such as (1-p1)×(1-p2)×(1-p3)) buffers the impact of parameter fluctuations. This improves the model's accuracy, enhances its robustness, and enables it to adapt to process changes (such as fluctuations in polishing slurry parameters in CMP processes), reducing the risk of false alarms and false negatives.

[0035] 2. Standardized calculation rules (such as H2 product scoring) are used to encode expert knowledge into reusable data models. New engineers can get started without extensive experience: for example, component maintenance scores are automatically generated based on preset conditions, and engineers only need to input maintenance records to obtain health values. This directly addresses the problem in existing technologies where "new engineers' work efficiency is long-term limited due to experience barriers," simplifies operational processes, improves response speed (such as early warning of potential faults), and reduces training costs.

[0036] 3. The dynamic weighting mechanism (w1×H1 + w2×H2 + w3×H3) allows for adjusting weights according to different needs (e.g., emphasizing the yield dimension w1 to address high defect risk), enabling flexible switching of analysis perspectives. This supports multi-scenario applications (e.g., different parameter settings for CMP processes), improves model versatility, and facilitates the integration of new dimensions (e.g., adding other FDC parameters).

[0037] 4. By quantifying health risks in real time (e.g., issuing warnings when the overall score falls below a threshold), unplanned downtime can be reduced. For example, a decrease in the H3 score based on the FDC alarm rate (e.g., temperature fluctuations > ±2℃) can trigger maintenance, preventing defect accumulation that leads to yield decline (e.g., scratch defects in CMP processes). Ultimately, this ensures process consistency, improves wafer yield, and reduces operating costs (e.g., reducing scrap rates).

[0038] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention.

[0040] Figure 2 This is a system module diagram of Embodiment 2 of the present invention.

[0041] Figure 3 This is a system module diagram of Embodiment 3 of the present invention. Detailed Implementation

[0042] Terminology Explanation:

[0043] Chamber: A closed device used for wafer processing in semiconductor manufacturing; its health directly affects the processing yield.

[0044] FDC: Fault Detection and Classification; a system dimension for monitoring abnormal process parameters, including alarm rates for pressure / temperature / vacuum fluctuations.

[0045] Recipe-FDC: Fault Detection and Classification (FDC) system based on specific process recipes.

[0046] CMP: Chemical Mechanical Polishing; a wafer surface polishing process, used as a typical application scenario for cavity health monitoring.

[0047] True defect: A real physical defect that actually exists on the surface of a wafer.

[0048] Repeater defect: A defect that occurs repeatedly; a type of defect that occurs frequently.

[0049] Big defect: a significant defect in a wafer.

[0050] Spatial signature defect: a spatial characteristic defect, including scratches / ring-shaped defects.

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0052] Example 1

[0053] like Figure 1 As shown, a method for analyzing cavity health based on big data statistics includes the following steps:

[0054] Historical and real-time data of the cavity are collected, including yield dimension characteristics, component maintenance dimension, and FDC data dimension.

[0055] Based on the aforementioned yield dimension features, calculate the yield dimension health score. The yield dimension features include the number of True defects, Repeater defects, Big defects, and Spatial signature defects, calculated as follows: for each type of defect... Perform normalization to obtain the historical maximum value. and historical minimum Then the health value of a single defect is Then, the average of the four defect health values ​​is obtained. ;

[0056] Calculate the component maintenance health level based on the aforementioned component maintenance dimensions. The component maintenance dimensions include maintenance frequency, most recent maintenance interval, number of failures, and average maintenance cycle. The calculation method is as follows: Each component's health score is determined according to preset rules. ,in These represent the health scores of different components;

[0057] Calculate the health status of the FDC data dimension based on the aforementioned FDC data dimension. The FDC data dimensions include the standard deviation of pressure fluctuations, the alarm rate for exceeding temperature fluctuation limits, and the alarm rate for large vacuum fluctuations. The calculation method is as follows ;

[0058] Based on weight ,in Calculate the overall health score of the cavity. .

[0059] It should be noted that the parallel calculation of the yield (H1), component maintenance (H2), and FDC (Fulfilled Directional Control) dimensions forcibly integrates the originally independent dimensions. For example, in the CMP process scenario: the yield dimension H1 calculation is based on defect normalization (such as the number of True defects) to capture the impact of surface defects; the component maintenance dimension H2 calculation uses rule-based scoring (such as conditions based on maintenance frequency and failure count) to quantify mechanical wear risk; the FDC dimension H3 calculation is based on complementary alarm rates (such as the standard deviation of pressure fluctuations) to monitor process stability. The data from these three dimensions interact dynamically through weighted fusion (S = w1×H1 + w2×H2 + w3×H3) to form a holistic model. For example, a decrease in the H2 score (indicating component failure) may be linked to an increase in H1 defects, thereby achieving a breakthrough in the "inability to capture the synergistic effect of yield decline and component wear" mentioned in the background technology, and accurately quantifying systemic risks (such as predicting unplanned downtime).

[0060] Subjectivity is eliminated through big data statistical rules (such as historical data baseline calibration and alarm rate calculation): In the parts maintenance dimension, H2 scoring is based on preset objective conditions (such as a score of 0.9 for maintenance frequency matching the average cycle). These rules directly transform expert knowledge into executable algorithms, avoiding human error. Specifically, the judgment conditions for this dimension (such as health scores of 0.9, 0.7, etc.) are encoded into standardized logic, see [link to relevant documentation]. Figure 1As shown: H2's scoring rules (such as getting 0.9 points for "matching maintenance frequency and having 0 failures") ensure consistent scoring. New engineers only need to input data to get reliable output, solving the problem of "experience barriers limiting work efficiency".

[0061] Meanwhile, the H3 calculation ((1-p1)×(1-p2)×(1-p3)) for the FDC dimension and the H1 averaging for the yield dimension are both based on historical data statistics (such as a six-month or one-year cycle), improving the model's generalization ability. This directly improves accuracy and enhances robustness (such as buffering random fluctuations).

[0062] The weighted fusion mechanism (w1+w2+w3=1) allows users to customize weights based on the scenario (e.g., increasing w3 to enhance FDC monitoring), thus overcoming the limitation of "fixed analysis perspective." For example, in CMP processes, engineers can increase w2 (part dimension) to address high-frequency maintenance needs, achieving "flexible adjustment of analysis perspective." This not only improves response speed (e.g., real-time weight adjustment to adapt to process changes) but also reduces operating costs (e.g., reducing unnecessary maintenance). Ultimately, this helps engineers provide early warnings (e.g., comprehensive score warning thresholds), improving yield and ensuring production continuity.

[0063] Through the above approach, a cross-dimensional dynamic fusion model was established, replacing experience-based reliance with data-driven methods to achieve highly accurate and robust chamber health analysis, while empowering new engineers to operate efficiently. Overall, the innovative aspects of this invention (such as multi-dimensional comprehensive analysis and weight adjustment) translate into practical industrial benefits, driving progress in the semiconductor manufacturing field.

[0064] During implementation, technicians first collect data from sensors or factory systems (such as the FDC system). When calculating H1 for the yield dimension, for each type of defect (e.g., True defect), the maximum (max) and minimum (min) of historical data (e.g., the past year) are obtained. Assuming the current defect quantity x is 50, historical min = 0, and max = 100, the health value for a single defect is calculated as 1 minus (x min) divided by (max min), i.e., 1 - (50-0) / (100-0) = 0.5. The average of the four health values ​​(True defect, Repeaterdefect, Big defect, and Spatial signature defect) is used to obtain H1. When calculating H2 for the component maintenance dimension, preset rules are referenced: for example, if a component's maintenance frequency matches the average cycle, the most recent interval is within limits, and the number of failures is 0, then the score is 0.9. Assuming the scores for three components are 0.9, 0.7, and 0.9 respectively, then H2 = 0.9 × 0.7 × 0.9 = 0.567. When calculating H3 using FDC data, the alarm rates for pressure fluctuations, temperature fluctuations, and vacuum fluctuations are taken. Assuming alarm rates p1=0.1 (10%), p2=0.2, and p3=0.05, then H3 = (1-0.1) × (1-0.2) × (1-0.05) = 0.9 × 0.8 × 0.95 ≈ 0.684. Finally, weights are set (e.g., w1=0.4, w2=0.3, w3=0.3), and S = 0.4×H1 + 0.3×H2 + 0.3×H3 is calculated. In CMP processes, engineers can run this method in real time: obtain data through an API interface, perform calculations on the server, and output the S value for decision-making.

[0065] In this embodiment, the Spatial signature defect includes scratch defects or ring defects, whose health value is calculated independently of other defects, and the normalized weight of scratch defects is 1.2 times that of other defects.

[0066] The above approach enhances the sensitivity to critical defects (such as scratches, which can easily damage the wafer surface in CMP processes) and improves the accuracy of health assessment. By adjusting the weights, it better reflects actual process risks (Document 1 points out that scratch defects affect subsequent processes) and avoids general defects from masking serious problems.

[0067] During implementation, technicians process the Spatial Signature Defect (SSD) separately when calculating H1. For example, assuming the number of SSD defects is x=30, the historical minimum is 0, and the maximum is 50, the health value is calculated as 1 - (30-0) / (50-0) = 0.4. Since the weight is 1.2, this value is multiplied by 1.2 before averaging to get 0.48. Health values ​​for other defects (such as True defects) are used directly (without weighting). In CMP scenarios, SSD defects are often caused by polishing pad wear. Independent weighting allows for earlier warnings of high risks: for example, a low SSD health value in H1 calculation significantly lowers the average score, triggering maintenance.

[0068] In this embodiment, the normalization process uses the Z-score standardization method, and the single defect health value is calculated as follows: ,in This is the historical average. The standard deviation is the historical value.

[0069] The above approach improves the robustness of data processing. The Z-score method considers the data distribution (mean μ and standard deviation σ) and reduces the impact of extreme values. Compared with min-max normalization, it is more adaptable to the process environment with large fluctuations and ensures that the health values ​​are stable and reliable.

[0070] During implementation, technicians calculate the mean μ and standard deviation σ of historical defect data. For example, for historical data of True defects, μ=40 and σ=10; assuming the current x=50, the health value = 1 - (50-40) / 10 = 1 - 1 = 0 (negative values ​​can be clamped to 0 or adjusted). In CMP processes, the number of defects often follows a normal distribution, and Z-score can better handle abnormal batches: if a batch of defects suddenly increases to 60 (far exceeding μ), the health value is calculated as 1 - (60-40) / 10 = 1 - 2 = -1. After clamping to 0, the H1 average score reflects a severe anomaly.

[0071] In this embodiment, the weight Based on real-time alarm rate dynamic adjustment, the calculation method is as follows: ,in For the first Current alarm rate for the dimension.

[0072] The above scheme achieves adaptive weighting, improving the flexibility of the method; when the risk of a certain dimension is high (such as a surge in component failure rate), its weight is automatically increased to ensure that the comprehensive score S is more in line with the real-time state, thus solving the problem that fixed weights cannot respond to sudden changes.

[0073] During implementation, after server monitoring, the weights are promptly increased and displayed to engineers for processing. For example, assuming the alarm rates for yield dimension r1=0.2, component dimension r2=0.3, and FDC dimension r3=0.1, then the weights w1 = 0.2 / (0.2+0.3+0.1) ≈0.333, w2 = 0.3 / 0.6 ≈ 0.5, and w3 = 0.1 / 0.6 ≈ 0.167. In CMP processes, if the pressure fluctuation alarm rate suddenly increases (r3=0.4), the system automatically increases the weight of w3, making S more sensitive to FDC issues, allowing engineers to prioritize their handling.

[0074] This embodiment also includes the step of: when the overall health score is... When the value drops below 0.5, a maintenance warning will be automatically triggered and a shutdown command will be generated.

[0075] The above solution enables proactive early warning and reduces unplanned downtime; the threshold mechanism (S<0.5) quantifies risks based on data, avoids human error, and improves production continuity and yield.

[0076] During implementation, the system calculates the S value in real time. If S = 0.4 (below 0.5), it automatically sends a warning signal to the maintenance terminal (such as email or work order system) and generates a shutdown command. For example, in CMP equipment, when S decreases due to low component score H2, the system shuts down, and engineers immediately inspect the support platform to prevent the defect from expanding.

[0077] In this embodiment, the alarm threshold of the FDC data dimension is user-defined, including pressure fluctuation threshold (±0.3psi) or temperature fluctuation range (±2℃).

[0078] The above approach enhances the applicability of the method, allowing users to set thresholds based on equipment type (such as different semiconductor processes) to avoid a one-size-fits-all approach; custom thresholds align with actual risk tolerance, improving model accuracy.

[0079] During implementation, engineers input thresholds via a user interface; for example, in a CMP process, a pressure fluctuation threshold might be set to ±0.3 psi—an alarm is triggered if the standard deviation of pressure exceeds 0.3 psi for 10 consecutive batches. The alarm rate is calculated based on this custom threshold. For instance, a user might set a temperature threshold of ±1°C for high-temperature sensitive equipment; the system would then calculate p2 and output H3 accordingly.

[0080] In this embodiment, the method is applied to multiple cavities and the factory-level average health score is calculated. ,in For the first Health score for each cavity.

[0081] The above solutions are extended to factory-level monitoring, providing macro-level health insights; average scores help managers identify overall risk trends (such as aging of multiple devices), optimize resource allocation, and improve overall plant efficiency.

[0082] During implementation, the system independently runs the cavity health analysis method for each cavity. Assuming the factory has N=5 cavities with S values ​​of 0.7, 0.6, 0.8, 0.5, and 0.9 respectively, the average score S_avg = (0.7+0.6+0.8+0.5+0.9) / 5 = 0.7. In semiconductor factories, this score is used for monthly reports: if S_avg < 0.6, a factory-wide maintenance plan is initiated.

[0083] In this embodiment, the data acquisition step includes acquiring data in real time from a sensor or FDC system via an API interface, and filling missing values ​​with linear interpolation.

[0084] The above solution ensures data integrity and reduces evaluation bias caused by missing data; the API interface enables automated data collection, and linear interpolation handles brief interruptions, improving the reliability and real-time performance of the method.

[0085] In implementation, the system connects to the sensor network via API. If a batch of data is missing (e.g., temperature values), linear interpolation is used: assuming the previous batch value is 20℃ and the next batch value is 22℃, the missing value is interpolated to 21℃. In CMP equipment, this method avoids H3 calculation errors due to sensor malfunctions.

[0086] Example 2

[0087] like Figure 2 As shown, a cavity health analysis system is characterized by comprising:

[0088] The data acquisition module is used to acquire the yield dimension characteristics, component maintenance dimension, and FDC data dimension of the cavity in real time;

[0089] The health score calculation module includes:

[0090] The yield analysis unit is configured to calculate the number of four types of defects: True defect, Repeater defect, Big defect, and Spatial signature defect. ; Number of defects The historical maximum value, Number of defects The historical minimum value;

[0091] The component analysis unit is configured to output based on preset rules. ;in These represent the health scores of different components;

[0092] The FDC analysis unit is configured to calculate based on the alarm rate. ;in, These are the alarm rates for pressure fluctuation standard deviation, temperature fluctuation exceeding limits, and large vacuum fluctuation, respectively.

[0093] The comprehensive evaluation module is configured to be weight-based. Calculate the overall health score of the cavity ;

[0094] The early warning execution module is configured to... Maintenance commands are triggered at certain times.

[0095] The system disclosed in Example 2 implements the analysis method of Example 1 into hardware, achieving end-to-end automation, significantly improving efficiency and reducing reliance on manual labor. Its modular design facilitates integration into existing factory systems (such as MES or PLC control systems), reducing deployment time and costs. Real-time monitoring and automatic early warning functions help engineers intervene in potential faults in advance, avoiding unplanned downtime. For example, in a semiconductor plant, this system can automatically process massive amounts of data, allowing new engineers to operate it without experience, improving workflow consistency.

[0096] When implementing this system, technicians first deploy the hardware components. The data acquisition module connects to the cavity sensors and factory database via Ethernet or an industrial bus (such as Modbus) to acquire data in real time. For example, in a CMP process cavity, this module acquires yield-related data (such as the number of true defects) from polishing pressure sensors and defect detection cameras, component maintenance data (such as the maintenance frequency of wafer chucks) from the maintenance log system, and FDC data (such as real-time values ​​of pressure fluctuations) from the FDC system. The data acquisition frequency can be set to per minute or per batch to ensure real-time performance. The health calculation module is deployed on an embedded controller or industrial PC: When the yield analysis unit calculates H1, it assumes that the current number of True defects x=50, Repeater defects x=30, Big defects x=20, and Spatial signature defects x=10; historical data max=100, min=0, then the health value of a single defect = 1 - (x-min) / (max-min), such as the health value of a True defect = 1 - (50-0) / (100-0) = 0.5; the average of the four health values ​​is H1=0.35. The component analysis unit outputs the score of each component based on a preset rule table (such as the maintenance rules in document 1): for example, if the maintenance frequency of the wafer carrier matches the average cycle and the number of failures = 0, then score1=0.9; the number of failures of the polishing pad dresser = 2 times / half a year, then score2=0.5; H2 is calculated as score1×score2×... When the FDC analysis unit calculates H3, it takes the pressure fluctuation alarm rate p1=0.1, temperature fluctuation p2=0.2, and vacuum fluctuation p3=0.05, then H3=(1-0.1)×(1-0.2)×(1-0.05)=0.684. The comprehensive evaluation module runs on the same controller, with weights set w1=0.4, w2=0.3, and w3=0.3, calculating S=0.4×H1 + 0.3×H2 + 0.3×H3; assuming S=0.4 (below 0.5), the early warning execution module automatically sends a shutdown command to the PLC system, triggering an alarm light or SMS notification to the maintenance team. In the CMP production line, engineers can quickly deploy: after connecting the sensors, the system runs automatically without programming knowledge; when the S value decreases, the system responds immediately to prevent the expansion of polishing defects.

[0097] Example 3

[0098] like Figure 3 As shown, a cavity health analysis software system is characterized by running on a server cluster and includes:

[0099] The data interface unit is used to receive yield, component maintenance, and FDC data from the sensor network;

[0100] Process the engine and perform the following operations:

[0101] For the yield data, calculate and output the normalized health value. The yield dimension features include the number of True defects, Repeater defects, Big defects, and Spatial signature defects, calculated as follows: for each type of defect... Perform normalization to obtain the historical maximum value. and historical minimum Then the health value of a single defect is Then, the average of the four defect health values ​​is obtained. ;

[0102] Maintenance data for parts is generated based on a preset rule table. The component maintenance dimensions include maintenance frequency, most recent maintenance interval, number of malfunctions, and average maintenance cycle. The calculation method is as follows: Each component's health score is determined according to preset rules. ,in These represent the health scores of different components;

[0103] Calculate the product output for FDC alarm rate data. The FDC data dimensions include the standard deviation of pressure fluctuations, the alarm rate for exceeding temperature fluctuation limits, and the alarm rate for large vacuum fluctuations. The calculation method is as follows ;

[0104] Based on dynamic weight formula Calculate the overall score of the cavity ;

[0105] Control command unit, when When the value falls below the threshold, a shutdown signal is sent to the maintenance terminal.

[0106] The software system disclosed in this embodiment provides a lightweight and scalable solution that runs on a cloud or local server cluster, supporting large-scale cavity monitoring (such as an entire wafer fab) and reducing hardware costs. The processing engine efficiently processes real-time data streams, ensuring rapid response (millisecond-level latency). Dynamic weighting and automated control enhance flexibility, and engineers can customize parameters (such as alarm thresholds) through a web interface for remote management. This overcomes the deployment limitations of hardware systems, is suitable for multi-factory scenarios, improves resource utilization, and reduces maintenance workload.

[0107] When implementing this software system, technicians deploy it on a server cluster (such as a Kubernetes-based cloud environment). The data interface unit connects to the sensor network via APIs (such as RESTful interfaces): for example, in a CMP process, this unit receives real-time data streams from an IoT gateway, such as yield dimension data (via a defect detection system API), component data (from a maintenance database API), and FDC data (from a process control system API); if data is missing (e.g., a batch temperature value is missing), linear interpolation is used to fill the missing value, for example: assuming the previous batch value = 20℃ and the next batch value = 22℃, then the missing value = 21℃, ensuring calculation continuity. The processing engine runs on the server: first, it calculates H1 for the yield data using a normalization method, assuming the number of spatial signature defects x = 15, historical max = 50, min = 0, and health value = 1 - (15-0) / (50-0) = 0.7; then, it outputs H1 after weighted averaging. Secondly, H2 is generated based on the rule table: the rule table stores preset conditions (such as the maintenance score rules in document 1). Assuming that the maintenance frequency of the component "grinding fluid nozzle" is slightly higher than the average cycle, the score is read from the table as 0.7, and H2 is calculated as the product of the scores of all components. Thirdly, H3 is calculated for the FDC alarm rate: a user-defined threshold (such as pressure fluctuation ±0.3psi) is taken, and the alarm rate is calculated and then multiplied for output. Finally, the processing engine applies a dynamic weight formula (such as weight 4): assuming the current yield dimension alarm rate r1=0.2, component r2=0.3, and FDC r3=0.4, then the weights w1=0.2 / (0.2+0.3+0.4)=0.222, w2=0.333, w3=0.444, and S=w1×H1 + w2×H2 + w3×H3 is calculated. The control instruction unit monitors the S value: if S = 0.4 (below the threshold of 0.5), it automatically sends a shutdown signal to the maintenance terminal via a message queue (such as Kafka), such as via SMS or email. When implemented in a semiconductor plant, engineers only need to install the software on the existing server cluster and configure the API connection; after setting parameters through the web interface, the system continues to run. For example, when pressure fluctuations cause S to drop, the software immediately triggers a shutdown, and engineers can remotely intervene to adjust the process.

[0108] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for analyzing cavity health based on big data statistics, characterized in that, Includes the following steps: Historical and real-time data of the cavity are collected. The data includes yield dimension features, component maintenance dimension and FDC data dimension. The cavity is used in the chemical mechanical polishing process in semiconductor manufacturing. The data is acquired in real time from the sensor or FDC system through the API interface, and missing values ​​are filled by linear interpolation. Based on the aforementioned yield dimension features, calculate the yield dimension health score. The yield dimension features include the number of True defects, Repeater defects, Big defects, and Spatial signature defects, calculated as follows: for each type of defect... Perform normalization to obtain the historical maximum value. and historical minimum Then the health value of a single defect is Furthermore, the health value calculation for scratch defects within the Spatial Signature defect is independent of other defects, with its normalized weight being 1.2 times that of other defects. The average of the health values ​​for the four defects is then obtained. ; Calculate the component maintenance health level based on the aforementioned component maintenance dimensions. The component maintenance dimensions include maintenance frequency, most recent maintenance interval, number of failures, and average maintenance cycle. The calculation method is as follows: Each component's health score is determined according to preset rules. ,in These represent the health scores of different components; Calculate the health status of the FDC data dimension based on the aforementioned FDC data dimension. The FDC data dimensions include the standard deviation of pressure fluctuations, the alarm rate for exceeding temperature fluctuation limits, and the alarm rate for large vacuum fluctuations. The calculation method is as follows ; Based on weight ,in Calculate the overall health score of the cavity. The weights Based on real-time alarm rate dynamic adjustment, the calculation method is as follows: ,in For the first The current alarm rate of each dimension, and when the alarm rate of any dimension exceeds a preset threshold, the weight of that dimension is automatically increased.

2. The method according to claim 1, characterized in that, The spatial signature defect includes ring-shaped defects.

3. The method according to claim 1, characterized in that, It also includes the step of: when the overall health score is... When the value drops below 0.5, a maintenance warning will be automatically triggered and a shutdown command will be generated.

4. The method according to claim 1, characterized in that, The alarm thresholds for the FDC data dimension are user-defined, including pressure fluctuation thresholds or temperature fluctuation ranges.

5. The method according to claim 1, characterized in that, The method is applied to multiple cavities and calculates the factory-level average health score. ,in For the first Health score for each cavity.

6. A cavity health analysis system, characterized in that, include: The data acquisition module is used to acquire the yield dimension characteristics, component maintenance dimension and FDC data dimension of the cavity in real time. The cavity is used in the chemical mechanical polishing process in semiconductor manufacturing. The data is received in real time through the API interface and missing values ​​are filled by linear interpolation. The health score calculation module includes: The yield analysis unit is configured to calculate the number of four types of defects: True defect, Repeater defect, Big defect, and Spatialsignature defect. ; Number of defects The historical maximum value, Number of defects The historical minimum value, the health value of a single defect. Furthermore, the health value calculation for scratch defects within the Spatial Signature Defect category is independent of other defects, with a normalization weight that is 1.2 times that of other defects. The component analysis unit is configured to output based on preset rules. ;in These represent the health scores of different components; The FDC analysis unit is configured to calculate based on the alarm rate. ;in, These are the alarm rates for pressure fluctuation standard deviation, temperature fluctuation exceeding limits, and large vacuum fluctuation, respectively. The comprehensive evaluation module is configured to be weight-based. Calculate the overall health score of the cavity The weights Based on real-time alarm rate dynamic adjustment, the calculation method is as follows: ,in For the first The current alarm rate of each dimension, and when the alarm rate of any dimension exceeds a preset threshold, the weight of that dimension is automatically increased; The early warning execution module is configured to... Maintenance commands are triggered at certain times.

7. A cavity health analysis software system, characterized in that, Running on a server cluster, including: The data interface unit is used to receive yield, component maintenance and FDC data from the sensor network. The cavity is used in the chemical mechanical polishing process in semiconductor manufacturing. The data is acquired in real time from the sensor or FDC system through the API interface and missing values ​​are filled by linear interpolation. Process the engine and perform the following operations: For the yield data, calculate and output the normalized health value. The yield dimension features include the number of True defects, Repeater defects, Big defects, and Spatial signature defects, calculated as follows: for each type of defect... Perform normalization to obtain the historical maximum value. and historical minimum Then the health value of a single defect is Furthermore, the health value calculation for scratch defects within the Spatial Signature defect is independent of other defects, with its normalized weight being 1.2 times that of other defects. The average of the health values ​​for the four defects is then obtained. ; Maintenance data for parts is generated based on a preset rule table. The component maintenance dimensions include maintenance frequency, most recent maintenance interval, number of malfunctions, and average maintenance cycle. The calculation method is as follows: Each component's health score is determined according to preset rules. ,in These represent the health scores of different components; Calculate the product output for FDC alarm rate data. The FDC data dimensions include the standard deviation of pressure fluctuations, the alarm rate for exceeding temperature fluctuation limits, and the alarm rate for large vacuum fluctuations. The calculation method is as follows ; Based on dynamic weight formula Calculate the overall score of the cavity The weights Based on real-time alarm rate dynamic adjustment, the calculation method is as follows: ,in For the first The current alarm rate of each dimension, and when the alarm rate of any dimension exceeds a preset threshold, the weight of that dimension is automatically increased; Control command unit, when When the value falls below the threshold, a shutdown signal is sent to the maintenance terminal.

Images