An AI sequence modeling-based production abnormality tracing method and system

Abstract

The application discloses a production abnormality tracing method and system based on AI sequence modeling. The method obtains real-time process parameter information, environmental state information and production batch information in the production process, identifies the current product formula according to the production batch information and obtains the influence degree information of each process parameter on the production result, evaluates the operation history information of the key production equipment to infer the local state change information of the equipment, monitors the operation state of the environmental control equipment to identify the environmental disturbance information, calculates a comprehensive index reflecting the current production risk according to the process parameter deviation, the formula influence degree, the equipment local state change and the environmental disturbance information, judges the production abnormality according to the comprehensive index and identifies the main contributing situational factors, and generates a diagnosis report containing the situational factors and their interaction relationship. The application can realize the differentiation of abnormalities, early warning of complex defects and comprehensive diagnosis of root causes.

Description

Technical Field

[0001] This invention relates to the field of production anomaly tracing technology, and more specifically, to a production anomaly tracing method and system based on AI sequence modeling. Background Technology

[0002] In modern industrial production, especially in semiconductor wafer manufacturing, advanced production anomaly traceability systems are typically deployed to ensure product quality and production efficiency. The semiconductor wafer manufacturing process is extremely complex, involving hundreds of sequential process steps, each subject to strict control by hundreds of parameters. To ensure the consistency of quality and performance of each batch of wafers, factories deploy tens of thousands of sensors to collect detailed process information from each stage in real time, forming a vast, high-dimensional process information sequence database.

[0003] To effectively manage and quickly respond to potential anomalies in production, the foundry introduced a production anomaly tracing system based on artificial intelligence sequence modeling. Before deployment, this system established a benchmark capable of accurately characterizing standard process paths by learning from the process information sequences of millions of historically successful wafer production batches. This benchmark not only captures the parameter range of each process step but also learns the evolution and interrelationships of parameters over time. In the early stages of system operation, its performance was satisfactory. For production fluctuations caused by single equipment failures or obvious operational errors, the system could accurately and quickly locate the fault point, significantly shortening troubleshooting time and reducing scrap rates. However, with the market's increasing demands for chip performance and power consumption, the foundry decided to launch a low-power version of an existing chip. While this new product is architecturally similar to the original, achieving lower power consumption requires a series of fine-tuning adjustments to its manufacturing process, including reducing the thickness of certain key dielectric layers, fine-tuning the etching gas ratio, and changing the ion implantation dosage. These adjustments did not completely change the process flow but rather involved localized and targeted modifications to the existing standard process path. Once these new process formulations were put into production, the AI-based traceability system, built on data from older products, began issuing frequent false alarms, misinterpreting these new but expected changes in process parameters as continuous deviations from the old baseline path. To avoid frequent production line interruptions, engineers had to increase the alarm thresholds, but this made the system less sensitive to small, potentially real fluctuations.

[0004] Meanwhile, more insidious problems on the production line began to slowly emerge. In one of the multiple plasma etching chambers used for etching critical metal layers, very slow, localized wear or material deposition began to appear on key internal components (such as gas distribution showerheads or RF electrodes). This manifested as subtle non-uniformities in plasma density or energy distribution across the wafer surface, undetectable by conventional monitoring systems. Furthermore, the cleanroom environmental control system experienced intermittent malfunctions; air handling unit dampers occasionally jammed, causing minute, instantaneous fluctuations in localized temperature and humidity—fluctuations below the overall environmental alarm threshold. When all these seemingly independent, minute factors converged at a specific moment, a non-linear amplification effect occurred, ultimately leading to specific types of defects on the wafers. These defects were often only discovered later in the production process, by which time the entire batch of wafers might already be unusable. Summary of the Invention

[0005] This invention provides a production anomaly tracing method and system based on AI sequence modeling, which at least solves the problems of existing AI-based production anomaly tracing systems in the face of complex production scenarios such as product formula iteration, equipment implicit degradation and instantaneous environmental disturbances. These systems cannot distinguish between expected process adjustments and real anomalies, and cannot identify the nonlinear superposition effect of multiple small factors, resulting in false alarms and missed alarms and a lack of in-depth causal relationship diagnosis.

[0006] In a first aspect, the present invention provides a method for tracing the source of production anomalies based on AI sequence modeling, comprising the following steps: Acquire real-time process parameter information, environmental status information, and production batch information during the production process, and perform time synchronization and data preparation on the information. Based on the production batch information, identify the product formula used in the current production batch, and obtain information on the degree of influence of each process parameter under the product formula on the production results; Based on the production batch information, assess the operating history information of key production equipment, and infer the local state change information of the equipment based on the operating history information; Based on the environmental status information, monitor the operating status of the environmental control equipment, identify the differences between the environmental control commands and the actual execution, and generate environmental disturbance information. Associate the environmental disturbance information with the production batch information. Based on the deviation of the process parameters, the degree of influence of the product formula on the parameters, the local state change information of the equipment, and the environmental disturbance information, calculate a comprehensive index reflecting the current production risk. Based on the comprehensive index, production anomalies are determined, and the situational factors that have the main impact on the anomalies are identified, generating a diagnostic report that includes the situational factors and their interactions.

[0007] Secondly, this invention provides a production anomaly tracing system based on AI sequence modeling, the system comprising: The information acquisition and preparation module is used to acquire real-time process parameter information, environmental status information, and production batch information during the production process, and to perform time synchronization and data preparation of the information. The formula identification and impact analysis module is used to identify the product formula used in the current production batch based on the production batch information, and to obtain information on the degree of influence of each process parameter under the product formula on the production results. The equipment assessment and status inference module is used to assess the operating history information of key production equipment based on the production batch information, and infer the local status change information of the equipment based on the operating history information. The environmental monitoring and risk calculation module is used to monitor the operating status of environmental control equipment based on the environmental status information, identify the differences between environmental control commands and actual execution and generate environmental disturbance information, associate the environmental disturbance information with the production batch information, and calculate a comprehensive index reflecting the current production risk based on the deviation of the process parameters, the degree of influence of the product formula on the parameters, the local state change information of the equipment and the environmental disturbance information. The anomaly diagnosis and tracing module is used to determine production anomalies based on the comprehensive index, identify the situational factors that have the main impact on the anomalies, and generate a diagnostic report containing the situational factors and their interactions.

[0008] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect above.

[0009] Compared with the prior art, the present invention has the following beneficial technical effects: By constructing a multi-dimensional situational awareness risk index that incorporates product formulation sensitivity, equipment degradation status, environmental disturbance impact, and parameter deviation characteristics, intelligent comprehensive evaluation of production anomalies is achieved. This index can distinguish between anticipated process adjustments introduced by new product formulations and genuine minor equipment degradation and environmental disturbances, avoiding false alarms while maintaining sensitivity to real anomalies. Furthermore, by introducing resonance amplification coefficients and synergistic / antagonistic effect analysis, the nonlinear amplification effect resulting from the superposition of multiple minor factors at a specific moment can be identified, enabling in-depth diagnosis of the root causes of complex defects.

[0010] Details of one or more embodiments of the present invention are set forth in the following drawings and description, so that other features, objects and advantages of the invention will be more readily understood. Attached Figure Description

[0011] Figure 1This is a flowchart of the production anomaly tracing method based on AI sequence modeling in Example 1.

[0012] Figure 2 This is a flowchart of step S4 in Example 1, which calculates a comprehensive index reflecting the current production risk.

[0013] Figure 3 This is a flowchart of step S3 in Example 1, inferring the local state change information of the device.

[0014] Figure 4 This is another flowchart for calculating the comprehensive index reflecting the current production risk in step S4 of Example 1.

[0015] Figure 5 This is a flowchart of step S32 in Example 1, which identifies repetitive signal patterns that do not conform to known degradation patterns.

[0016] Figure 6 This is a flowchart of step S2 in Example 1, which involves obtaining information on the degree of influence of each process parameter on the production results.

[0017] Figure 7 This is a block diagram of the production anomaly tracing system based on AI sequence modeling in Example 2. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0019] Example 1

[0020] Embodiment 1 of the present invention provides a method for tracing the source of production anomalies based on AI sequence modeling. Figure 1 This is a flowchart illustrating a production anomaly tracing method based on AI sequence modeling, according to an exemplary embodiment. Figure 1 As shown, the method includes the following steps: S1, acquire real-time process parameter information, environmental status information and production batch information during the production process, and perform time synchronization and data preparation on the information; In this step, the system connects to the equipment automation programs and manufacturing execution systems on the production line via Industrial Ethernet protocols, commonly including OPC UA and MQTT. Sensor data, including temperature, pressure, gas flow rate, RF power, and wafer stage rotation speed within the reaction chamber, is collected from the equipment controller, along with production management information such as wafer batch number, product recipe identifier, and process start and end times. Environmental monitoring data is transmitted in real-time to the central data server via an independent wireless sensor network, which can use LoRaWAN or Zigbee protocols. All data is accompanied by a high-precision timestamp and synchronized with a network time protocol server. The received raw data undergoes format conversion and preliminary verification. Missing sensor readings are filled using linear interpolation, and outliers significantly exceeding physical limits are replaced using median filtering or moving averages.

[0021] S2, based on the production batch information, identify the product formula used in the current production batch, and obtain information on the degree of influence of each process parameter under the product formula on the production results; In this step, when a wafer batch enters a specific process equipment via an automated material handling system, the manufacturing execution system sends a production instruction for that batch to the traceability unit, which includes a unique product recipe identifier used by the current wafer. The traceability unit receives this identifier and uses it as the recipe context input for the current process step. The system maintains a structured recipe sensitivity configuration table, stored in a relational database, predefined by process engineers during the product recipe design phase based on process physics principles, materials science knowledge, and expected product performance. The table contains a list of key process parameters corresponding to each recipe identifier, as well as the sensitivity level of each parameter under that recipe. When the system identifies a recipe identifier, it queries and loads the sensitivity information of all key parameters under the corresponding recipe from this configuration table, providing an accurate contextual basis for subsequent real-time risk assessment.

[0022] S3, based on the production batch information, evaluate the operating history information of key production equipment, and infer the local state change information of the equipment based on the operating history information; In this step, information such as equipment uptime, number of wafers processed, last maintenance date, and replacement records of critical components are obtained from the Manufacturing Execution System (MES) and the Computerized Maintenance Management System (CMS). Critical components, such as the gas distribution showerhead and RF electrodes within the etching chamber, have this information stored in the equipment's digital archive. Long-term trend analysis is performed on key sensor data: for example, for the RF reflected power of the etching chamber, its moving average over a specific number of days is calculated. If this value shows a consistently small but steady upward trend relative to the initial stable operation phase after the equipment is new or following a major overhaul, it is inferred that the RF matching network or electrodes may be experiencing initial degradation. For the valve position feedback signal of the gas flow controller, a slight delay in the time required to reach the set value is monitored. If the response time consistently exceeds a certain proportion of the historical normal value, it is inferred that the valve may be showing initial signs of mechanical wear or jamming.

[0023] S4. Based on the environmental status information, monitor the operating status of the environmental control equipment, identify the difference between the environmental control command and the actual execution, and generate environmental disturbance information. Associate the environmental disturbance information with the production batch information. Based on the deviation of the process parameters, the degree of influence of the product formula on the parameters, the local state change information of the equipment, and the environmental disturbance information, calculate a comprehensive index reflecting the current production risk. In this step, real-time data is acquired from the controller of the cleanroom air handling unit via an industrial communication interface. This includes set opening commands and actual position feedback for damper valves, and set and actual speed feedback for fan motors. Control commands and actual feedback are continuously compared. If a discrepancy is detected, a local environmental disturbance event marker is immediately generated, recording its occurrence time, duration, affected area, and inferred disturbance type. When evaluating the current process step, it is checked whether there are any local environmental disturbance event markers related to the area where the process equipment is located. For each real-time acquired process parameter data point or its changing trend, a situational awareness risk index is calculated in real time. This comprehensive index... The calculation formula is:

[0024] in, The parameter deviation magnitude weight is calculated by weighting the absolute deviation between the real-time value of the current process parameter and the median value of the parameter within the expected normal range under the current formulation. Obtained through function mapping, the greater the deviation, the higher the weight; The duration weight of parameter deviation is obtained by mapping the duration of parameter deviation from the expected range through a logarithmic function; The formula sensitivity coefficient is determined based on the formula sensitivity information loaded in step S2. The weight of the equipment degradation state is determined based on the local state change information of the equipment evaluated in step S3. The weights for environmental disturbances are assigned based on whether there are related local environmental disturbance events in the current process steps. Specifically, the parameter deviation magnitude is mapped to the weight value via a nonlinear function, the parameter deviation duration is mapped via a logarithmic function, the formulation sensitivity coefficient directly amplifies the risk based on a predefined sensitivity level, the equipment degradation state weight increases when the equipment is marked as being in an initial degradation state, and the environmental disturbance impact weight increases when related environmental disturbance events exist. The initial values ​​of each weight and coefficient are preset by process engineers based on experience and experimental data, and can be fine-tuned during operation.

[0025] S5. Based on the comprehensive index, determine the production anomaly, identify the situational factors that have the main impact on the anomaly, and generate a diagnostic report containing the situational factors and their interactions. In this step, a configurable risk threshold is set. When the situation-aware risk index of any process parameter exceeds this threshold, an anomaly alarm is immediately triggered. After triggering, the components leading to the high-risk index are analyzed, identifying the situational factors that contribute most to the final risk index during the calculation process. A traceability report with structured text and visual charts is generated, clearly explaining how the degradation of specific equipment and the superposition of specific environmental disturbances jointly led to the process anomaly under the current product formulation's sensitivity requirements for process parameters. Historical trend charts of relevant parameters, links to equipment maintenance records, and environmental event logs are provided for further analysis by process engineers. The report helps engineers understand the underlying causes of the anomaly, rather than just superficial parameter deviations, enabling them to take more accurate corrective measures.

[0026] In the technical solution of the above embodiments, by acquiring real-time process parameters, environmental conditions, and production batch information during the production process, and loading sensitivity information of process parameters according to the product formula, the operating history of key production equipment is evaluated to infer local state changes. The operating status of environmental control equipment is monitored to identify environmental disturbances. Parameter deviation characteristics, formula sensitivity, equipment degradation status, and the impact of environmental disturbances are comprehensively incorporated into the calculation of the context-aware risk index. Based on the risk index, production anomalies are judged, and the main contributing contextual factors are identified to generate a diagnostic report. This method breaks through the traditional fixed threshold judgment mode, realizing intelligent comprehensive risk assessment based on production context. It can distinguish between expected process adjustments introduced by new product formulas and actual anomalies, and can also identify the nonlinear superposition effect of multiple minor factors, thereby providing a diagnosis of deep causal relationships.

[0027] In some preferred embodiments, it is assumed that a semiconductor wafer fab is producing a new type of low-power logic chip. This chip has significantly higher requirements for plasma uniformity during the etching process than previous products. When a batch of wafers using the new formulation enters the etching chamber for critical metal layer etching, the system acquires real-time process parameters, environmental status information, and production batch information in step S1, completing time synchronization and data preparation. Step S2 identifies the formulation identifier of the new formulation and loads the extremely high sensitivity coefficients set for RF power stability and gas flow stability for this formulation. Step S3, by evaluating the operating history of the chamber, detects a continuous slight upward trend in RF reflected power over the past two weeks, inferring initial wear in the gas distribution shower head and assigning a degradation state weight to the equipment. Step S4 detects a brief jamming of the air handling unit's damper valve during the critical etching stage, generating a transient environmental disturbance event marker and assigning an environmental disturbance impact weight. Simultaneously, a small fluctuation in RF power, deviating from the expected range of the new formulation by ten percent, is detected, lasting approximately ten seconds. The context-aware risk index multiplies the weights of parameter deviation magnitude, parameter deviation duration, formulation sensitivity coefficient, equipment degradation status, and environmental disturbance impact to obtain a comprehensive index value exceeding a preset risk threshold. Step S5 triggers an alarm and generates a traceability report based on this comprehensive index, clearly indicating that under the current high-sensitivity formulation, initial wear of the shower head and localized airflow instability in the cleanroom jointly contribute to the risk of uneven etching. Engineers prioritize checking the shower head status and adjust the damper maintenance plan based on the report, preventing wafer scrapping before large-scale defects occur.

[0028] In this embodiment, Figure 2 This is a flowchart illustrating the calculation of a comprehensive index reflecting current production risks in step S4 of Example 1. (Refer to...) Figure 2 The step S4, which calculates a comprehensive index reflecting the current production risk, includes: S41 performs real-time spectrum analysis on the raw sensor signals of key production equipment, extracts the instantaneous spectrum characteristics of the signal within a short time window, and identifies hidden high-frequency local oscillation modes. In this step, the raw sensor signals of key process equipment are sampled at high frequency. Signal types include raw voltage or current signals of radio frequency power, raw cavity pressure signals, and raw light intensity signals of plasma emission spectra. The sampling rate is at least twice the expected highest oscillation frequency. The formula for calculating the Short-Time Fourier Transform (STFT) is:

[0029] in The original signal, For sliding window functions, For a moment, The frequency is used to obtain the energy distribution of the signal at different times and frequencies through this transformation. Particular attention is paid to specific frequency ranges in the spectrum related to initial wear or material deposition of equipment components. By analyzing the energy intensity, peak frequency drift, or bandwidth variations within these frequency ranges, hidden high-frequency local oscillation modes are identified. If the energy intensity at a specific frequency remains consistently higher than the background noise level within a continuous short-time Fourier transform window, and its variation pattern matches the predictions of a physical model of known component wear, it is marked as a high-frequency oscillation event, and its frequency, intensity, and duration are recorded.

[0030] S42 monitors the internal control loops of cleanroom environmental control equipment, compares control commands with actual feedback signals from actuators in real time, and identifies transient environmental disturbance events. In this step, the set opening commands and actual position feedback of the damper valves, as well as the set and actual speed feedback of the fan motor, are read from the air handling unit controller at a high frequency. The instantaneous difference between the commands and feedback is continuously calculated. If this difference exceeds a preset small threshold within a short time window and lasts for an extremely short time, it is identified as an instantaneous damper jamming perturbation. Further analysis of the instantaneous rate of change and recovery characteristics of this difference classifies it into specific environmental perturbation signatures, such as rapid instantaneous jamming or slow-recovery jamming. These environmental perturbation event markers, with precise timestamps and affected areas, can capture short-term, non-periodic local environmental fluctuations that traditional environmental sensors may miss due to insufficient sampling rate or data granularity.

[0031] S43, when the high-frequency local oscillation mode and the instantaneous environmental disturbance event are detected simultaneously, resonance analysis is initiated to analyze whether there is time synchronization, frequency alignment or phase correlation between the high-frequency local oscillation mode and the instantaneous environmental disturbance event, and to calculate the resonance amplification factor based on the resonance intensity, duration and the sensitivity of the product formula used in the current production batch to resonance. In this step, the resonance analysis module checks whether the precise occurrence time window of the high-frequency oscillation event overlaps with the duration window of the environmental perturbation event. If overlap exists, further analysis is performed: during the occurrence of the environmental perturbation, does the center frequency of the high-frequency oscillation experience a slight drift, bringing it closer to a certain inherent resonant frequency of the plasma, or does the bandwidth of the oscillation momentarily narrow, indicating energy concentration? The module also checks whether the energy intensity of the high-frequency oscillation exhibits a momentary, unexpected nonlinear increase during the occurrence of the environmental perturbation. For example, if the energy intensity of the high-frequency oscillation increases by a specific percentage momentarily during the environmental perturbation, and neither the oscillation alone nor the environmental perturbation would cause such a large increase, it is determined to be resonance amplification. Based on the detected resonance intensity and duration, and the sensitivity of the current product formulation to resonance, the resonance amplification factor is calculated. The formula for calculating the resonance amplification factor is:

[0032] in, The frequency alignment strength is a value between 0 and 1, representing the degree of frequency alignment. The energy amplification ratio represents the percentage increase in detected instantaneous energy. The formulation resonance sensitivity is a preset sensitivity coefficient defined by process engineers based on experience and experimental data. This coefficient reflects the strength of the nonlinear amplification effect produced by the superposition of two minute factors, enabling the system to quantify this hidden resonance destructive force.

[0033] S44, The resonance amplification coefficient is incorporated into the calculation of the comprehensive index to form an enhanced risk assessment, so as to obtain the comprehensive index reflecting the current production risk; In this step, the calculated resonance amplification factor is directly multiplied into the original context-aware risk index calculation to form an enhanced risk assessment model:

[0034] in, As the weight for the magnitude of parameter deviation, The duration of parameter deviation is the weight, and S_recipe is the recipe sensitivity coefficient. Weights for equipment degradation states. Weighting for the impact of environmental disturbances. This is the resonance amplification factor. When the enhanced risk index exceeds a preset variable risk threshold, it is considered a production anomaly. The enhanced risk index is calculated by multiplying the original weights for parameter deviation amplitude, parameter deviation duration, formulation sensitivity coefficient, equipment degradation status, and environmental disturbance impact by an additional resonance amplification factor. When the enhanced risk index exceeds the preset variable risk threshold, it is considered a production anomaly. When generating the traceability report, priority is given to analyzing whether the resonance amplification factor is significantly greater than the benchmark value. If this factor contributes significantly, the report explicitly states that high-frequency local oscillations in the equipment resonate with instantaneous environmental disturbances, leading to a non-linear amplification of process risk. The report also details the characteristics of the high-frequency oscillations involved, the type of environmental disturbance, and the sensitivity of the current formulation to the resonance.

[0035] In the technical solution of the above embodiments, high-frequency spectrum analysis is performed on the original sensor signals of key production equipment to identify concealed high-frequency local oscillation modes. Real-time comparison is performed on the internal control loops of cleanroom environmental control equipment to identify instantaneous environmental disturbances. When both occur simultaneously, resonance analysis is initiated to calculate the resonance amplification coefficient, which is then integrated into the calculation of the situation-aware risk index to form an enhanced risk assessment. This solution solves the problem that traditional monitoring methods cannot capture concealed equipment degradation and short-term environmental disturbances due to insufficient signal sampling or inadequate data granularity. Furthermore, by quantifying the resonance amplification effect, it achieves accurate identification and early warning of complex process anomalies caused by the nonlinear superposition of multiple minute factors.

[0036] In some preferred embodiments, it is assumed that the wear of the shower head inside the etching cavity evolves into a more subtle physical effect, inducing localized high-frequency electromagnetic oscillations in the plasma. This oscillation frequency is approximately 120 Hz, close to the 100 Hz sampling rate of the main RF power sensor, and is filtered out as high-frequency noise during conventional data preprocessing, thus failing to be captured by the conventional parameter deviation detection mechanism in step S4. Simultaneously, the damper jamming in the cleanroom air handling unit evolves into intermittent, non-periodic short-term jamming, each lasting only two to three seconds. In this case, step S41 successfully identifies the subtle 120 Hz localized high-frequency oscillation mode by high-frequency sampling of the raw voltage signal from the RF power sensor and extracting instantaneous spectral features within an extremely short time window using short-time Fourier transform. Step S42 identifies the two-to-three-second instantaneous damper jamming disturbance event by high-frequency comparison of the damper valve's control command with its actual position feedback. In step S43, the resonance analysis module detected a momentary nonlinear increase in the energy intensity of the 120 Hz oscillation during the damper jamming period, and the oscillation frequency drifted towards the plasma's natural resonant frequency, which was determined to be a resonance amplification effect. The resonance amplification coefficient was calculated. In step S44, this coefficient was incorporated into the risk index calculation, causing the enhanced risk index to exceed the threshold and trigger an alarm. The source tracing report indicated that the high-frequency local oscillation of the equipment resonated with the instantaneous environmental disturbance, leading to a nonlinear amplification of process risks. This helped engineers understand the complex causal relationship between latent degradation and the superposition of instantaneous disturbances, and led to the development of improvement measures to increase the original signal sampling rate and the resonance detection frequency.

[0037] In this embodiment, Figure 3 This is a flowchart illustrating the inference of local state change information of the device in step S3 of Embodiment 1. (Refer to...) Figure 3 The step of inferring the local state change information of the device in step S3 includes: S31 acquires real-time data of multiple operating parameters from key production equipment; In this step, real-time data of multiple operating parameters are acquired from the controllers of key production equipment (such as etching chambers and thin-film deposition chambers) via the industrial Ethernet protocol. These parameters include RF power, reflected power, chamber pressure, gas flow rate, wafer stage temperature, and valve position feedback. This data is acquired at a high frequency to ensure that subtle changes in the equipment's operating status can be captured.

[0038] S32, perform interaction pattern analysis on the real-time data of the multiple operating parameters to identify changes in the correlation strength between the parameters; In this step, correlation and cointegration analyses are performed on multiple collected operating parameters to identify typical correlation patterns between parameters under normal operating conditions. By continuously tracking changes in the correlation coefficients or cointegration relationships between parameters, when the correlation strength of a parameter pair deviates from its historical normal range, it is marked as an anomaly. Changes in this interaction pattern between parameters often reveal changes in the internal physical state of the equipment earlier than absolute shifts in a single parameter. For example, a slight shift in the normal ratio between RF power and reflected power may indicate changes in the matching network or electrode state.

[0039] S33, perform background feature analysis on the real-time data of the multiple operating parameters to identify subtle shifts in the signal background features; In this step, the statistical characteristics of each operating parameter signal are continuously monitored, including background features such as the signal's mean, variance, skewness, kurtosis, and spectral distribution. A dynamic baseline model is established for each parameter, and the difference between the current signal background characteristics and historical baselines is compared using a sliding window approach. When a signal background characteristic shows a sustained, subtle shift beyond the normal fluctuation range, it is marked as a potential anomaly, even if the parameter's amplitude remains within the normal alarm threshold. For example, although the valve position signal of a gas flow controller may still fluctuate within the normal range, a continuous change in the skewness of its probability distribution may indicate initial mechanical wear of the valve.

[0040] S34, based on the changes in the correlation strength between the parameters and the subtle shifts in the signal background features, infer the local state change information of the device; In this step, the analysis results of the correlation strength changes between parameters and the signal background feature shifts are fused, and the local state change information of the equipment is inferred through a preset equipment degradation mode mapping rule. The mapping rule is jointly established by process engineers and equipment engineers based on the equipment physical model and historical maintenance experience, and specifies the most likely degradation type corresponding to a specific combination of correlation strength change patterns and background feature shifts.

[0041] In the technical solution of the above embodiments, by acquiring real-time data of multiple operating parameters, performing interaction pattern analysis between parameters and signal background feature analysis, the local state change information of the device is comprehensively inferred from two dimensions: changes in correlation strength and background feature shift. This method does not rely on the absolute threshold judgment of a single parameter, but extracts early signs of device degradation from the evolution of relationships between parameters and subtle changes in signal statistical characteristics, achieving earlier detection and more accurate assessment of latent local degradation of the device.

[0042] In this embodiment, Figure 4 This is another flowchart illustrating the calculation of a comprehensive index reflecting current production risks in step S4 of Example 1. (Refer to...) Figure 4 The step S4, which calculates the comprehensive index reflecting the current production risk, further includes: S45, Real-time acquisition of information on deviations of process parameters, the degree of influence of product formula on parameters, local state changes of equipment, and environmental disturbances; In this step, complete contextual information for the current process step is collected in real time from various analysis modules. This includes the direction and magnitude of the deviation of the current process parameters from the expected range of the formulation, the sensitivity coefficients of the current formulation to each parameter loaded in step S2, the inferred local state changes of the equipment and their confidence levels in step S3, and the environmental disturbance event markers, durations, and affected areas identified from the AHU controller comparison. This information is integrated into a unified contextual description structure for use in subsequent synergistic / antagonistic effect analysis.

[0043] S46. Based on the collected information, analyze whether there is a synergistic or antagonistic effect among the various situational factors. The synergistic effect is manifested in the fact that when multiple situational factors exist simultaneously, their impact on production risk is greater than the sum of the individual effects of each factor. The antagonistic effect is manifested in the fact that when multiple situational factors exist simultaneously, their impact on production risk is less than the sum of the individual effects of each factor. In this step, the interaction effects of three situational factors—parameter sensitivity of the product formulation, localized equipment degradation, and environmental disturbance events—are analyzed in pairs and in triplicate. A multivariate analysis of variance (ANOVA) model is established to analyze the interaction effects of these three situational factors. For each process parameter, a linear model including main effects and interaction effects is established:

[0044] in, As a formulation sensitivity factor, As a factor of equipment degradation, Environmental disturbance factor, interaction term coefficient Reflecting synergistic or antagonistic effects, The random error term represents the unexplained variance in the model beyond the main effects and interaction effects. A synergistic effect is indicated when the interaction term coefficient is positive and statistically significant; an antagonistic effect is indicated when the interaction term coefficient is negative and statistically significant. Typical manifestations of synergistic effects include: when highly sensitive formulations are combined with early equipment degradation, even small parameter fluctuations can easily trigger a chain reaction; when equipment degradation is combined with local environmental disturbances, environmental instability accelerates the spread of adverse effects from degraded components. Typical manifestations of antagonistic effects include: the impact of certain types of equipment degradation is partially buffered under specific environmental conditions. The analysis results provide a basis for adjusting the weights of various situational factors in the risk index calculation.

[0045] S47. Based on the analysis results of the synergistic or antagonistic effects, adjust the weights or contributions of each situational factor in the calculation of the comprehensive index, and apply the adjusted weights or contributions to the calculation of the comprehensive index to obtain the comprehensive index reflecting the current production risk. In this step, the original weights of each situational factor are dynamically adjusted based on the analysis results of synergistic or antagonistic effects. When a synergistic effect is detected, the combination of interacting factors involved will be given an additional synergistic amplification factor, making the risk index more sensitive to the combination of related factors. When an antagonistic effect is detected, the weight contribution of the combination of related factors is appropriately reduced to avoid over-alarming. The adjusted weights are substituted into the multiplicative model of the situation-aware risk index to calculate the final comprehensive index, making the risk assessment more accurately reflect the true impact of multi-factor interactions.

[0046] In the technical solution of the above embodiments, by further analyzing the synergistic or antagonistic effects among various situational factors when calculating the comprehensive index, and dynamically adjusting the weights or contributions of each factor based on the interaction effect analysis results, a refined modeling of the interactive effects of multiple factors is achieved. This method can distinguish between simple superposition and nonlinear interaction, avoids overestimating or underestimating the effects of independent factors, and makes the risk assessment results more consistent with the complex physical processes in actual production.

[0047] In this embodiment, Figure 5 This is a flowchart of step S32 in Example 1, which identifies repetitive signal patterns that do not conform to known degradation patterns. (Refer to...) Figure 5 When inferring local state change information of the device based on the aforementioned operational history information, the method further includes: S321, acquire raw operating status signals of key production equipment, and analyze the raw signals to identify repetitive signal patterns that do not match known degradation patterns; In this step, raw operating status signals are acquired from key production equipment using a high-frequency sampling rate. Signal types include raw voltage signals of radio frequency power and raw analog signals of cavity pressure. Pattern recognition algorithms are used to analyze the signals for recurring patterns, employing methods such as unsupervised anomaly detection based on autoencoders or similarity matching based on dynamic time warping. These patterns are then compared with known degradation patterns recorded in a database, filtering out recurring signal patterns whose match with all known patterns is below a preset threshold. These patterns may indicate new types of degradation or abnormal behavior in the equipment.

[0048] S322, combining the product formula information of the current production batch, analyze the interaction mechanism between material properties and equipment components; This step involves obtaining information related to process materials from the product formulation used in the current production batch, including the types of precursor gases, etching gas ratios, and the composition of the deposited materials. Based on materials science and equipment physics models, the potential physical or chemical interaction mechanisms between these materials and key components of the equipment, such as shower head materials and cavity coatings, under specific process conditions like high temperatures and plasma environments, are analyzed. For example, certain fluorinated etching gases may cause slightly abnormal corrosion to alumina shower heads under specific power conditions.

[0049] S323, associate the repetitive signal pattern with the material properties and the device interaction mechanism to establish preliminary association rules, and track the device operation performance and wafer defect status under the preliminary association rules; In this step, based on the repetitive signal patterns identified in step S321 and the material-device interaction mechanism analyzed in step S322, a preliminary correlation rule between the two is established through time series alignment and causal inference. Subsequently, the subsequent defect detection results of the corresponding batch of wafers when the correlation rule is triggered are continuously tracked to accumulate evidence to strengthen or correct the confidence of the correlation rule.

[0050] S324, confirm that the repetitive signal pattern is a new device local state change pattern, and establish a mapping with the local state change information; In this step, when the tracking data shows a statistically significant correlation between the repetitive signal pattern and a specific type of wafer defect, and the influence of other known factors has been ruled out, the signal pattern is identified as a new local state change pattern of the equipment. This new pattern, along with its corresponding equipment components, degradation mechanisms, and process influence characteristics, is then entered into the equipment degradation pattern database to establish a mapping relationship between the pattern and local state change information.

[0051] In the technical solution of the above embodiments, pattern recognition is performed on the original signals of equipment operation status to discover repetitive signals that do not conform to known degradation patterns. The interaction mechanism of equipment components is analyzed in conjunction with the material characteristics of the product formulation, association rules are established and tracked for verification, and finally, new local state change patterns of the equipment are confirmed and a mapping is established. This method endows the system with the ability to learn on its own and discover new degradation patterns, enabling it to adapt to unprecedented degradation behaviors of equipment components under new product formulations and process conditions, thus improving the comprehensiveness and accuracy of state inference.

[0052] In some preferred embodiments, it is assumed that during long-term operation, a critical etching chamber exhibits novel localized degradation on its radio frequency (RF) electrodes that does not match any known degradation patterns in the database. Step S321 acquires the raw voltage signal of the chamber's RF power and analyzes its signal characteristics using a pattern recognition algorithm. A repetitive signal pattern is identified that does not match any known showerhead uniform wear pattern or electrode corrosion pattern. This pattern is characterized by intermittent pulses of energy at a specific frequency during specific phases of each process run. Step S322, combined with the product formulation containing the specific novel etching gas used in the current production batch, analyzes that the gas composition, under specific RF power conditions, may produce unconventional localized chemical corrosion on the coating material on the electrode surface. Step S323 correlates the repetitive signal pattern with the material-device interaction mechanism, establishing a preliminary correlation rule between the signal pattern and coating corrosion. Defect detection in the corresponding batch of wafers is continuously tracked over several weeks when this rule is triggered. Tracking data indicates that these batches of wafers exhibit consistent micro-defects in specific areas. After excluding the influence of other known factors, step S324 identifies the repetitive signal pattern as a new local state change pattern of the equipment, indicating local corrosion of the coating. A mapping relationship between this pattern and local state change information is established and entered into the equipment degradation pattern database for subsequent state inference and risk assessment of similar equipment.

[0053] In this embodiment, Figure 6 This is a flowchart illustrating the process of obtaining information on the degree of influence of each process parameter on the production result in step S2 of Example 1. (Refer to...) Figure 6 The step S2, which involves obtaining information on the degree of influence of each process parameter on the production results, includes: S21, Based on the production batch information, identify the product formula used in the current production batch, and obtain the initial setting range of each process parameter under the product formula; In this step, the product formula identifier for the current production batch is obtained from the Manufacturing Execution System (MES). The initial setting ranges for each process parameter under this formula are then retrieved by querying the formula management database. These ranges include the target values, upper and lower limits, and normal fluctuation bands for each parameter. These initial setting ranges are defined by process engineers during the product formula development phase and reflect the process conditions required by the current formula. The obtained setting ranges provide a benchmark for subsequent assessment of whether the actual operating data of the parameters deviates from expectations.

[0054] S22 monitors the actual operating data of each process parameter and the corresponding wafer production result data in real time during the production process; In this step, as the wafer batches enter each process step for processing, real-time operational data for each process parameter is acquired at high frequency. This includes real-time values ​​of parameters such as gas flow rate, chamber pressure, RF power, and temperature, as well as their time-varying curves. Simultaneously, early quality proxy index data closely related to this process step are also acquired. For example, in the thin film deposition step, instantaneous film thickness uniformity and refractive index distribution are obtained using an online ellipsometer; in the etching step, specific spectral line intensity changes are obtained using plasma emission spectroscopy; and instantaneous temperature distribution is obtained using a wafer surface temperature distribution sensor. Furthermore, wafer production result data related to subsequent process steps or inspection steps, such as optical defect detection results and electrical test data, are also collected.

[0055] S23, when a new product formula is detected to be put into production, the parameter impact assessment process is initiated to perform correlation analysis on the actual operating data and the wafer production result data to identify the relationship between the process parameters and the production results; In this step, when the Manufacturing Execution System (MES) reports that a new product formulation or a minor derivative of an existing formulation has been put into production that does not perfectly match the existing formulation identifier, the Parameter Impact Assessment (DAI) process is automatically initiated. During this process, correlation analysis is performed on the collected actual operating data and wafer production result data to calculate the dynamic correlation strength between real-time changes in each process parameter and changes in early quality proxy indicators. The formula for calculating the Dynamic Correlation Strength Index (DAI) is as follows:

[0056] in, The Pearson correlation coefficient is the relationship between the instantaneous rate of change of process parameters and the instantaneous rate of change of surrogate indicators. This is a nonlinear factor based on the degree of nonlinear mapping between the two, obtained by fitting a polynomial regression model and evaluating the residuals. This represents the average lag time of the proxy indicator response relative to parameter fluctuations. A higher value indicates a more significant impact of parameter fluctuations on early quality proxy indicators. The dynamic correlation strength index comprehensively considers the correlation coefficient, nonlinearity factor, and lag time, reflecting the sensitivity of the current batch to specific parameters in real time.

[0057] S24, Based on the aforementioned relationship of change, dynamically adjust the information on the degree of influence of each process parameter under the product formula on the production results; In this step, the continuously calculated dynamic correlation strength is compared with the preset sensitivity of the parameter in the current formulation sensitivity configuration table. If the dynamic correlation strength of a process parameter is significantly higher than the expected impact level implied by its corresponding sensitivity level in the current configuration table over multiple consecutive batches or a specific time period, then the parameter is determined to have implicit high sensitivity to the current product variant. The system will dynamically increase the formulation sensitivity coefficient of the parameter and provide a calibration confidence score.

[0058] In the technical solution of the above embodiments, by identifying the current product formulation and obtaining the initial set range, the actual operating data of process parameters and the corresponding wafer production result data are monitored in real time. When a new product formulation is introduced, a parameter impact assessment process is initiated to perform correlation analysis, and the impact degree information of each process parameter is dynamically adjusted according to the changing relationship. This method endows the system with the ability to adaptively calibrate parameter sensitivity based on real-time production data, enabling it to automatically identify implicitly high-sensitivity parameters caused by new materials or new structures when facing product iteration and generalized formulation identifiers, thereby avoiding the underreporting of key process anomalies due to inaccurate identification of context sensitivity.

[0059] In some preferred embodiments, it is assumed that the wafer fab quickly deployed the iterative version v2.0 after launching the low-power chip v1.0. Due to the tight development cycle, the engineer maintaining the recipe sensitivity configuration table only updated the sensitivity settings for core parameters directly related to the new structure, while retaining the medium sensitivity setting of v1.0 for the initial flow rate ramp-up rate parameter of the precursor gas in the thin film deposition chamber. Simultaneously, the manufacturing execution system unified multiple v2.0 derivative versions into a generalized recipe identifier after a software update. When a batch of wafers actually v2.0b enters the thin film deposition chamber, step S21 identifies the generalized recipe identifier, and the sensitivity of the initial flow rate ramp-up rate is set to medium in the sensitivity configuration loaded in step S22. Step S23, in real-time monitoring, detects a continuous strong correlation between the small fluctuations in this gas flow rate parameter and the instantaneous film thickness uniformity provided by the online ellipsometer. The calculated dynamic correlation strength index consistently falls within the high sensitivity range, far exceeding the expected level corresponding to medium sensitivity. Step S24 triggers a sensitivity calibration event, dynamically raising the formulation sensitivity coefficient of this parameter from a moderate level to an extremely high level, along with a calibration confidence score. When the gas flow controller experiences a slight hysteresis in response due to minor mechanical wear, the adaptively calibrated formulation sensitivity coefficient assesses this deviation as high risk and triggers an alert. The traceability report indicates that the system automatically detected the parameter's implicit high sensitivity to the current product variant by observing the dynamic correlation between parameter fluctuations and early quality proxy indicators, recommending that engineers update the configuration table and inspect the gas flow controller.

[0060] In this embodiment, when inferring the local state change information of the device based on the operating history information, the method of multi-resolution signal decomposition is also used to analyze the original signal of the device operating state: S35 acquires raw operating status signals from key production equipment, including broadband vibration, acoustic, and electrical signals; In this step, raw operational signals covering a wide frequency band are acquired by installing or utilizing existing sensors on key production equipment. These include vibration signals acquired by accelerometers, acoustic signals acquired by acoustic emission sensors, and electrical signals acquired by current and voltage sensors. These signals cover a wide frequency range, from low-frequency mechanical vibrations to high-frequency electrical fluctuations, and are acquired synchronously at a sampling rate at least twice the highest frequency of interest.

[0061] S36, Perform multi-resolution decomposition on the original signal of the operating state, and extract the instantaneous energy, frequency distribution and phase features of the signal at different time scales; In this step, multi-resolution signal decomposition methods such as wavelet transform or empirical mode decomposition are used to process the acquired raw signal. Multi-scale decomposition is performed by selecting appropriate wavelet basis functions; commonly used wavelet basis functions include Daubechies wavelet and Morlet wavelet. Detail and approximate components of the signal are extracted at different time scales, corresponding to different frequency ranges. For each scale component, the instantaneous energy envelope, instantaneous frequency distribution, and instantaneous phase difference or phase synchronization index between adjacent sensor signals are calculated. These multi-resolution features can reveal subtle changes in the device's operating status at different levels.

[0062] S37, based on the instantaneous energy, frequency distribution and phase characteristics, identify weak, non-periodic signal patterns related to the initial degradation of equipment components, wherein the amplitude of the signal patterns is lower than the conventional alarm threshold; In this step, statistical analysis is performed on the extracted multi-resolution features, focusing on weak anomaly patterns with amplitudes below the conventional alarm thresholds but persistently occurring within specific time-frequency regions. This involves detecting intermittent, pulse-like increases in instantaneous energy on fine time scales; examining the frequency distribution for new frequency components or frequency drift; and assessing the phase characteristics for a gradual loss of phase synchronization. Because these signal patterns are non-periodic, weak, and potentially correlated with specific degradation mechanisms, they are considered early signs of initial degradation in equipment components.

[0063] S38. Based on the evolution trend of the weak, non-periodic signal pattern, combined with the sensitivity of the current production batch's product formula to the degradation pattern, the local state change information of the equipment is inferred, and the future degradation trend of the equipment is predicted. In this step, time-series tracking is performed on the identified weak aperiodic signal patterns to analyze their evolution trends, including whether the frequency of the signal patterns is increasing, whether the energy intensity is continuously rising, and whether the time-frequency region of influence is expanding. Combined with information on the sensitivity of the current production batch's product formulation to equipment degradation patterns, comprehensive inferences are made regarding the current local state changes of the equipment, including the identification of degradation sites, the division of degradation stages, and the quantification of degradation severity. Furthermore, trend extrapolation or degradation models are used to predict the future degradation trend of the equipment, estimating at what production batch quantity or operating time the degradation will develop to a level that may lead to significant process defects, thereby providing data-driven decision support for the equipment's preventative maintenance plan.

[0064] In the technical solution of the above embodiments, by acquiring broadband vibration, acoustic, and electrical raw signals, multi-resolution decomposition is used to extract instantaneous energy, frequency distribution, and phase features at different time scales. This identifies weak, non-periodic signal patterns with amplitudes below conventional alarm thresholds, and combines this with product formulation sensitivity to infer local equipment status changes and predict future degradation trends. This method overcomes the limitations of traditional monitoring methods that only focus on whether signal amplitude exceeds standards and periodic fault modes. It can discover early, weak signs of equipment degradation from multi-dimensional, multi-scale signal features and assess their actual risk impact in conjunction with the production context, achieving earlier detection and more accurate prediction of latent equipment degradation.

[0065] In this embodiment, when inferring local state change information of the device based on the operation history information, a data-driven device evaluation method based on the operation history information is also included: S36a, Obtain the operating history information of key production equipment, including sensor data, maintenance records, component replacement information, and process parameters.

[0066] In this step, complete operational history information of key production equipment is extracted from the Manufacturing Execution System (MES), Computerized Maintenance Management System (CMS), and Supervisory Control and Data Acquisition (SCADA) system. Sensor data includes sampling sequences of parameters such as temperature, pressure, power, and flow rate during each process run; maintenance records include the time, content, and replacement parts list for each preventive and corrective maintenance; component replacement information includes the installation date, expected lifespan, and runtime of each key component; and process parameters include the specific process recipes and setpoints executed on the equipment for each batch. This multi-source, heterogeneous data is integrated into a unified equipment digital archive, forming a complete operational history view.

[0067] S36b, perform data cleaning and preprocessing on the running history information. The data cleaning and preprocessing includes missing value filling, outlier handling, and timestamp alignment to obtain preprocessed running history information.

[0068] In this step, the integrated operational history information undergoes systematic data cleaning and preprocessing. For missing value imputation, missing sampling points in the sensor data are filled using interpolation or estimation based on the changing trends of adjacent data. For outlier handling, statistical methods such as the interquartile range or three-standard-deviation method are used to identify outliers that significantly deviate from the normal range, and these outliers are marked or replaced. For timestamp alignment, data from different systems are aligned along a unified timeline to ensure accurate temporal correspondence between data sequences during subsequent feature extraction.

[0069] S36c, Feature extraction is performed on the preprocessed running history information, and the feature extraction includes statistical features, time domain features and frequency domain features.

[0070] In this step, multi-dimensional feature vectors are extracted from the preprocessed data. Statistical features include the mean, variance, skewness, kurtosis, and percentiles of each sensor parameter within a specific statistical window; time-domain features include autocorrelation coefficient, differential sequence statistics, and signal zero-crossing rate; frequency-domain features include the dominant frequency components, spectral energy distribution, and frequency band energy proportion obtained through Fourier transform or wavelet transform.

[0071] S36d, based on the aforementioned features, infer local state change information of the device, including wear, corrosion, deposition, or fatigue.

[0072] In this step, the extracted multidimensional feature vectors are input into a pre-trained machine learning model, which is trained based on historical degradation cases and maintenance verification data. The model infers the most likely local state change type of the current equipment by mapping the feature patterns to known degradation types, and provides the probability distribution and confidence score for each type. Local state change types include wear, corrosion, deposition, or fatigue. The inference results also include the approximate location of the degradation and a comparison with the historical degradation rate of similar equipment at similar operating stages.

[0073] In the technical solution of the above embodiments, complete historical operating information, including sensor data, maintenance records, component replacement information, and process parameters, is acquired. After systematic data cleaning and preprocessing, multi-dimensional feature vectors with statistical, time-domain, and frequency-domain features are extracted. Based on historical degradation cases, a model is trained to infer local state changes such as wear, corrosion, deposition, or fatigue. This method fully utilizes multi-source operating data throughout the equipment's entire lifecycle, achieving a systematic assessment of local state changes in the equipment through a data-driven approach.

[0074] In summary, the production anomaly tracing method based on AI sequence modeling provided in Embodiment 1 of this invention integrates product formula sensitivity, local equipment degradation status, the impact of instantaneous environmental disturbances, and parameter deviation characteristics into a context-aware risk index, achieving comprehensive risk assessment based on production context. Through high-frequency spectrum analysis and real-time comparison of control loops, it identifies hidden high-frequency local oscillations and instantaneous environmental disturbances, and quantifies the resonance amplification effect between them. By analyzing the synergistic or antagonistic effects between contextual factors, it refines the modeling of multi-factor interaction. Through multi-resolution decomposition of the original equipment signal, it captures weak non-periodic degradation signals below the conventional alarm threshold. Through adaptive calibration of parameter sensitivity based on real-time production feedback, it addresses the challenges posed by product iteration and generalized formula identifiers. By continuously discovering and confirming new equipment degradation patterns, it maintains comprehensive awareness of changes in equipment status. The synergistic effect of these features enables this method to distinguish anomalies, provide early warnings of complex defects, and comprehensively diagnose root causes in complex production contexts characterized by frequent product formula updates, hidden degradation of equipment components, and instantaneous environmental disturbances.

[0075] Example 2

[0076] Embodiment 2 of the present invention provides a production anomaly tracing system based on AI sequence modeling. Figure 7 This is a block diagram of a production anomaly tracing system based on AI sequence modeling, according to Example 2. Figure 7 As shown, the system includes: The information acquisition and preparation module 01 is used to acquire real-time process parameter information, environmental status information and production batch information during the production process, and to perform time synchronization and data preparation on the information.

[0077] The formula identification and impact analysis module 02 is used to identify the product formula used in the current production batch based on the production batch information, and to obtain information on the degree of influence of each process parameter under the product formula on the production results.

[0078] The equipment evaluation and status inference module 03 is used to evaluate the operating history information of key production equipment based on the production batch information, and infer the local status change information of the equipment based on the operating history information.

[0079] The environmental monitoring and risk calculation module 04 is used to monitor the operating status of environmental control equipment based on the environmental status information, identify the difference between environmental control commands and actual execution and generate environmental disturbance information, associate the environmental disturbance information with the production batch information, and calculate a comprehensive index reflecting the current production risk based on the deviation of the process parameters, the degree of influence of the product formula on the parameters, the local state change information of the equipment and the environmental disturbance information.

[0080] The anomaly diagnosis and tracing module 05 is used to determine production anomalies based on the comprehensive index, identify the situational factors that have a major impact on the anomalies, and generate a diagnostic report containing the situational factors and their interactions.

[0081] The production anomaly tracing system based on AI sequence modeling described in the above embodiment, through the collaborative work of five modules, acquires real-time process parameter information, environmental status information, and production batch information during the production process. It identifies and loads parameter influence information based on the product formula, assesses the operating history information of key production equipment to infer local state changes, monitors the operating status of environmental control equipment to identify environmental disturbances, comprehensively incorporates various situational factors into risk index calculation, judges production anomalies based on the risk index, identifies the main contributing factors, and generates a diagnostic report. This system achieves intelligent comprehensive evaluation of anomalies in complex production scenarios, distinguishing between expected process adjustments and actual anomalies, and identifying the nonlinear superposition effects of multiple minor factors, providing a diagnosis of deep causal relationships.

[0082] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method provided in Embodiment 1 above.

[0083] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for tracing the source of production anomalies based on AI sequence modeling, characterized in that, Includes the following steps: Acquire real-time process parameter information, environmental status information, and production batch information during the production process, and perform time synchronization and data preparation on the information. Based on the production batch information, identify the product formula used in the current production batch, and obtain information on the degree of influence of each process parameter under the product formula on the production results; Based on the production batch information, assess the operating history information of key production equipment, and infer the local state change information of the equipment based on the operating history information; Based on the environmental status information, monitor the operating status of the environmental control equipment, identify the differences between the environmental control commands and the actual execution, and generate environmental disturbance information. Associate the environmental disturbance information with the production batch information. Based on the deviation of the process parameters, the degree of influence of the product formula on the parameters, the local state change information of the equipment, and the environmental disturbance information, calculate a comprehensive index reflecting the current production risk. Based on the comprehensive index, production anomalies are determined, and the situational factors that have the main impact on the anomalies are identified, generating a diagnostic report that includes the situational factors and their interactions.

2. The method according to claim 1, characterized in that, The steps of monitoring the operating status of environmental control equipment based on the environmental status information, identifying the differences between environmental control commands and actual execution and generating environmental disturbance information, associating the environmental disturbance information with the production batch information, and calculating a comprehensive index reflecting the current production risk based on the deviation of process parameters, the degree of influence of the product formula on the parameters, the local state change information of the equipment, and the environmental disturbance information include: Real-time spectrum analysis of raw sensor signals from key production equipment is performed to extract instantaneous spectrum features of the signals within a short time window and identify hidden high-frequency local oscillation modes. Monitor the internal control loops of cleanroom environmental control equipment, compare control commands with actual feedback signals from actuators in real time, and identify transient environmental disturbance events; When the high-frequency local oscillation mode and the instantaneous environmental disturbance event are detected simultaneously, resonance analysis is initiated to analyze whether there is time synchronization, frequency alignment or phase correlation between the high-frequency local oscillation mode and the instantaneous environmental disturbance event, and to calculate the resonance amplification factor based on the resonance intensity, duration and the sensitivity of the product formula used in the current production batch to resonance. The resonance amplification factor is incorporated into the calculation of the comprehensive index to form an enhanced risk assessment, thereby obtaining the comprehensive index that reflects the current production risk.

3. The method according to claim 1, characterized in that, The step of evaluating the operating history information of key production equipment based on the production batch information, and inferring local state change information of the equipment based on the operating history information, includes: Obtain real-time data on multiple operating parameters from key production equipment; Interaction pattern analysis is performed on the real-time data of the multiple operating parameters to identify changes in the strength of the correlation between the parameters; Background feature analysis is performed on the real-time data of the multiple operating parameters to identify subtle shifts in the signal background features; Based on the changes in the correlation strength between the parameters and the subtle shifts in the signal background features, the local state change information of the device is inferred.

4. The method according to claim 1, characterized in that, The step of monitoring the operating status of environmental control equipment based on the environmental status information, identifying the difference between environmental control commands and actual execution and generating environmental disturbance information, associating the environmental disturbance information with the production batch information, and calculating a comprehensive index reflecting the current production risk based on the deviation of process parameters, the degree of influence of the product formula on the parameters, the local state change information of the equipment, and the environmental disturbance information, further includes: Real-time data collection includes deviations from process parameters, the degree of influence of product formulation on parameters, local state changes of equipment, and environmental disturbance information. Based on the collected information, it is analyzed whether there is a synergistic or antagonistic effect among the various situational factors. The synergistic effect is manifested in the fact that when multiple situational factors exist simultaneously, their impact on production risk is greater than the sum of the individual effects of each factor. The antagonistic effect is manifested in the fact that when multiple situational factors exist simultaneously, their impact on production risk is less than the sum of the individual effects of each factor. Based on the analysis results of the synergistic or antagonistic effects, the weights or contributions of each situational factor in the calculation of the comprehensive index are adjusted, and the adjusted weights or contributions are applied to the calculation of the comprehensive index to obtain the comprehensive index reflecting the current production risk.

5. The method according to claim 3, characterized in that, When inferring local state change information of the device based on the aforementioned operational history information, the method further includes: Acquire raw operating status signals of key production equipment and analyze the raw signals to identify repetitive signal patterns that do not conform to known degradation patterns; Based on the product formulation information of the current production batch, analyze the interaction mechanism between material properties and equipment components; The repetitive signal patterns are correlated with the material properties and the device interaction mechanism to establish preliminary correlation rules, and the device operation performance and wafer defect status under the preliminary correlation rules are tracked. The repetitive signal pattern is confirmed as a new local state change pattern of the device, and a mapping is established with the local state change information.

6. The method according to claim 3, characterized in that, When inferring local state change information of the device based on the aforementioned operational history information, the method further includes: Obtain raw operating status signals from key production equipment, including broadband vibration, acoustic, and electrical signals; The original signal of the operating state is decomposed into multi-resolution components to extract the instantaneous energy, frequency distribution, and phase features of the signal at different time scales. Based on the instantaneous energy, frequency distribution, and phase characteristics, weak, non-periodic signal patterns related to the initial degradation of equipment components are identified, and the amplitude of the signal patterns is lower than the conventional alarm threshold. Based on the evolution trend of the weak, non-periodic signal pattern, combined with the sensitivity of the current production batch's product formula to the degradation pattern, the local state change information of the equipment is inferred, and the future degradation trend of the equipment is predicted.

7. The method according to claim 1, characterized in that, The step of identifying the product formula used in the current production batch based on the production batch information, and obtaining information on the degree of influence of each process parameter under the product formula on the production result includes: Based on the production batch information, identify the product formula used in the current production batch and obtain the initial setting range of each process parameter under the product formula; During the production process, real-time monitoring of the actual operating data of each process parameter and the corresponding wafer production results data is conducted. When a new product formulation is detected to be put into production, a parameter impact assessment process is initiated to perform correlation analysis on the actual operating data and the wafer production result data to identify the relationship between the changes in the process parameters and the production results. Based on the aforementioned relationship of change, the influence of each process parameter under the product formula on the production results is dynamically adjusted.

8. The method according to claim 3, characterized in that, When inferring local state change information of the device based on the aforementioned operational history information, the method further includes: Obtain the operating history information of key production equipment, including sensor data, maintenance records, component replacement information, and process parameters; The historical data of the operation is cleaned and preprocessed, including missing value filling, outlier handling and timestamp alignment, to obtain the preprocessed historical data of the operation. Feature extraction is performed on the preprocessed running history information, including statistical features, time-domain features, and frequency-domain features; Based on the aforementioned characteristics, information on local state changes of the device is inferred, including wear, corrosion, deposition, or fatigue.

9. A production anomaly tracing system based on AI sequence modeling, characterized in that, The system includes: The information acquisition and preparation module is used to acquire real-time process parameter information, environmental status information, and production batch information during the production process, and to perform time synchronization and data preparation of the information. The formula identification and impact analysis module is used to identify the product formula used in the current production batch based on the production batch information, and to obtain information on the degree of influence of each process parameter under the product formula on the production results. The equipment assessment and status inference module is used to assess the operating history information of key production equipment based on the production batch information, and infer the local status change information of the equipment based on the operating history information. The environmental monitoring and risk calculation module is used to monitor the operating status of environmental control equipment based on the environmental status information, identify the differences between environmental control commands and actual execution and generate environmental disturbance information, associate the environmental disturbance information with the production batch information, and calculate a comprehensive index reflecting the current production risk based on the deviation of the process parameters, the degree of influence of the product formula on the parameters, the local state change information of the equipment and the environmental disturbance information. The anomaly diagnosis and tracing module is used to determine production anomalies based on the comprehensive index, identify the situational factors that have the main impact on the anomalies, and generate a diagnostic report containing the situational factors and their interactions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images