Machine sensor grading method and machine management and control method based on sensor data

Abstract

The present application relates to production equipment management and control technical field, particularly to a kind of machine sensor grading method and machine control method based on sensor data, machine sensor grading method includes: according to the data of each sensor in historical FDC data constructs its historical data Trace;Abnormality determination rule of data Trace is constructed, whether the real-time data Trace of each sensor is abnormal based on the abnormality determination rule constructed is judged, first determination result is obtained and is archived;The correlation between the first determination result of each sensor and corresponding product quality data is judged, and each sensor is graded based on correlation, and the corresponding level of each sensor is obtained.In the present application, each sensor is graded according to the sensor raw data in FDC data, the key information in raw data is not lost, the grading result is more accurate, and machine control is facilitated according to the grading result.

Description

Technical Field

[0001] This invention relates to the field of production equipment control technology, and in particular to a machine sensor classification method and a machine control method based on sensor data. Background Technology

[0002] Machine sensor data, as direct data for monitoring machines, reflects the current operating status (i.e., working status) of the machines in real time. When abnormalities occur in the machine's status, they directly impact product quality; therefore, the analysis and mining of machine sensor data is crucial. With the development of big data and advanced intelligent technologies, the storage and analysis of hundreds of billions of sensor data points has become possible. Consequently, major semiconductor manufacturers are no longer limited to analyzing product yield rates but aim to achieve early warning through the study of massive amounts of underlying machine data.

[0003] However, the data volume and data dimensions of the machine's sensors are enormous, and it is time-consuming and labor-intensive to manage them manually. As the complexity of semiconductor manufacturing processes increases year by year and the data that sensors can collect becomes larger and larger, engineers need to spend a lot of manual analysis work on anomaly monitoring and troubleshooting. Due to the huge number of analysis dimensions (which can be more than hundreds of thousands of parameters) and the differences in parameter characteristics, it is impossible for humans to achieve comprehensive and effective control of FDC data, and it is often impossible to detect abnormal machine conditions in a timely manner.

[0004] Furthermore, the current method of selecting key monitoring parameters and setting rules based on human experience suffers from incomplete monitoring and low accuracy. Engineers currently select some key parameters based on process experience to build FDC statistical models and set corresponding control rules. On the one hand, some parameters cannot comprehensively monitor machine status, and some abnormal states may go undetected due to the neglect of other parameters. On the other hand, using FDC statistics / feature data not only results in the loss of some information but may also lead to inappropriate rule control due to insufficient human experience, causing frequent false alarms. Therefore, it is necessary to propose a machine control method that can comprehensively monitor and accurately provide early warnings. Summary of the Invention

[0005] One of the objectives of this invention is to provide a machine sensor classification method and a machine management and control method based on sensor data. By accurately classifying the sensors and using the abnormal detection results of the sensors together with their classification results as the detection basis, the management and control effect of the machines can be greatly improved.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for classifying machine sensor, comprising the following steps:

[0007] Acquire historical FDC data of the machine;

[0008] Construct a historical data trace based on the data from each sensor in the historical FDC data;

[0009] Construct rules for anomaly detection in data traces;

[0010] Acquire real-time data from each sensor on the machine, and construct a real-time data trace for each sensor based on the acquired data;

[0011] Based on the constructed anomaly determination rules, determine whether the real-time data trace of each sensor is abnormal, obtain the first determination result of each sensor, and archive it;

[0012] When the number of archived first determination results reaches a predetermined number, the correlation between the first determination results of each sensor and the corresponding product quality data is determined, and each sensor is classified based on the correlation to obtain the corresponding level of each sensor.

[0013] Furthermore, the data trace includes non-vibration curves, and the anomaly determination rule includes a first anomaly determination rule corresponding to the non-vibration curves; the first anomaly determination rule is configured as follows:

[0014] The similarity between the non-vibration curve to be judged and each class within the corresponding dataset in the pre-constructed first baseline library is calculated. If the similarity with each class is less than the corresponding similarity threshold, the non-vibration curve to be judged is marked as abnormal; otherwise, the non-vibration curve to be judged is marked as normal.

[0015] Furthermore, the method for constructing the first baseline library includes the following steps:

[0016] Align the non-vibration curves formed by data collected by the same sensor at different times in the historical data trace according to the processing steps;

[0017] Using data from the same sensor under the same recipe of the same tool as a dataset, non-vibration curves in the historical data traces of each dataset are clustered based on a clustering algorithm, thereby obtaining one or more classes in each dataset; after removing classes identified as noise, a first baseline library is obtained in the form of datasets containing classes.

[0018] The similarity threshold is calculated using the following method:

[0019] Calculate the similarity of non-vibration curves within each class in each dataset of the first baseline library, and use the mean of the pairwise similarity between non-vibration curves within each class as the similarity threshold for that class.

[0020] Furthermore, the data trace includes a vibration curve, and the anomaly determination rule includes a second anomaly determination rule corresponding to the vibration curve, wherein the second anomaly determination rule is configured as follows:

[0021] The frequency of the vibration curve to be judged is calculated, and an anomaly detection algorithm is used to determine whether the frequency of the vibration curve is normal relative to the vibration curves in the second baseline library; wherein, the second baseline library is formed by vibration curves in the historical data trace of each sensor.

[0022] A sliding time window is introduced to calculate the local amplitude of the vibration curve to be judged within each time window. An anomaly detection algorithm is used to calculate whether the local amplitude of the vibration curve within each time window is normal relative to the vibration curve in the second baseline library.

[0023] If the frequency and local amplitude of the vibration curve to be judged are both normal within each time window, then the data of the vibration curve to be judged is marked as normal; otherwise, the data of the vibration curve to be judged is marked as abnormal.

[0024] Furthermore, in the step of determining the correlation between the first determination result of each sensor and the product quality data, and classifying each sensor based on the correlation to obtain the corresponding level of each sensor, a random forest model is used to determine the correlation and classify the sensors.

[0025] Furthermore, in the step of classifying each sensor based on the correlation, sensors whose first determination result is strongly correlated with product quality are classified as high-risk sensors, sensors whose first determination result is weakly correlated with product quality are classified as low-risk sensors, and sensors whose first determination result is unrelated to product quality are classified as no-risk sensors.

[0026] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is to provide a machine control method based on sensor data, comprising the following steps:

[0027] The sensor classification method described above is used to classify the sensors of the machine tool to obtain the corresponding level of each sensor.

[0028] Acquire real-time data from each sensor on the machine, and construct a real-time data trace for each sensor based on the acquired data;

[0029] Based on the constructed anomaly determination rules, it is determined whether the real-time data trace of each sensor is abnormal, and a second determination result is obtained;

[0030] The machine is controlled based on the corresponding level of each sensor and the second judgment result.

[0031] Furthermore, the step of controlling the machine based on the corresponding level of each sensor and the second determination result includes:

[0032] When the corresponding level of the sensor is a high-risk sensor and the second judgment result is N abnormalities, the machine is considered to be in an abnormal working state, and abnormal control is implemented on the machine; when the corresponding level of the sensor is a low-risk sensor and the second judgment result is M consecutive abnormalities, the machine is considered to be in an abnormal working state, and abnormal control is implemented on the machine; where M is greater than N, and N and M are both positive integers.

[0033] When the corresponding level of the sensor is a risk-free sensor, the second determination result of the risk-free sensor is checked and evaluated periodically. Based on the periodic evaluation results, it is determined whether the working status of the machine is abnormal, and corresponding control measures are implemented.

[0034] Furthermore, N takes the value 1.

[0035] Furthermore, when the second determination result is a data trace anomaly, the following steps are also performed:

[0036] Determine whether preventative maintenance was performed before the data trace anomaly occurred; if so, take the current time point as the cutoff point and obtain the FDC data of the relevant sensors before and after the multiple preventative maintenance time points prior to this point.

[0037] Based on the FDC data, calculate the fluctuation range of the FDC data of the relevant sensor before and after each preventive maintenance time node, and calculate the average fluctuation range of the FDC data of the relevant sensor before and after a single preventive maintenance.

[0038] The anomaly detection rules for data traces from relevant sensors are adjusted based on the average fluctuation amplitude.

[0039] In summary, the machine sensor classification method and machine control method of the present invention have the following unexpected technical effects:

[0040] (i) By systematically analyzing the correlation between abnormal detection results of machine FDC data and related quality data, this invention can more accurately find key sensors that may affect product quality, making sensor classification more accurate and providing a theoretical basis for subsequent machine management.

[0041] (ii) Classification is based on the sensor’s original data FDC Trace (sensor data trace) in the FDC data, which will not lose key information in the original data and makes the basis for classification more comprehensive.

[0042] (iii) The grading results are largely unrelated to the engineers’ human experience, greatly reducing the reliance on human intervention in the grading process and avoiding the impact of engineers’ lack of experience on the grading.

[0043] (iv) Based on the different degrees of correlation between sensor FDC data and product quality data, different risk levels are set for all sensors on the machine. Appropriate control rules are established based on historical data and abnormal situations. High-risk sensors are strictly controlled, and non-risk sensors are regularly evaluated. This can complement the current SPC system and greatly improve the quality control effect.

[0044] (v) Different detection methods are used for data traces in different scenarios and in different forms, so that the detection methods are more suitable for the corresponding scenarios or curve forms, are more targeted, can improve the accuracy of anomaly detection, and can carry out more effective and precise control, greatly reducing false alarms and false negatives. Attached Figure Description

[0045] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0046] Figure 1 This is a flowchart of an embodiment of the machine sensor classification method of the present invention.

[0047] Figure 2 This is a schematic diagram of three common anomalies in non-vibration curves. Among them, Figure 2 (a) is a schematic diagram of abnormal level values ​​in non-vibration curves. Figure 2 (b) is a schematic diagram of abnormal jump points in the non-vibration curve. Figure 2 (c) is a schematic diagram of the non-vibration curve Shift anomaly.

[0048] Figure 3 This is a schematic diagram illustrating two common anomalies in vibration curves. Among them, Figure 3 (a) is a schematic diagram of the abnormal frequency of the vibration curve. Figure 3 (b) is a schematic diagram of the abnormal amplitude of the vibration curve.

[0049] Figure 4 A schematic diagram illustrating the local amplitude within the time window used to calculate the vibration curve to be judged by introducing a sliding time window.

[0050] Figure 5 This is a schematic diagram illustrating the principle of the Random Forest algorithm.

[0051] Figure 6 This is a flowchart of an embodiment of the machine control method based on sensor data of the present invention. Detailed Implementation

[0052] The following disclosure provides various embodiments or examples for implementing different features of the invention. Specific examples of components and arrangements will be described below to simplify the invention. Of course, these are merely examples and are not intended to limit the invention. For example, in the following description, forming a first component above or on a second component may include embodiments where the first and second components are in direct contact, or embodiments where other components may be formed between the first and second components such that the first and second components are not in direct contact. Furthermore, reference numerals and / or characters may be repeated in various instances of the invention. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or configurations.

[0053] Furthermore, spatial relation terms such as "below," "under," "below," "above," and "above" may be used herein to readily describe the relationship between one element or component and another element (or component) or component (or component) as shown in the figure. In addition to the orientations shown in the figure, spatial relation terms will encompass various different orientations of the device in use or operation. The device may be positioned in other ways (rotated 90 degrees or in other orientations), and will be interpreted accordingly through the spatial relation descriptors used herein.

[0054] The above embodiments merely illustrate preferred implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention should be determined by the appended claims.

[0055] Please see Figure 1 , Figure 1 A flowchart of a method for determining the operating status of a machine tool based on sensor data is shown. In the illustrated embodiment, the method for determining the operating status of a machine tool based on sensor data includes the following steps:

[0056] S100. Acquire historical FDC data of the equipment. The data collected by the FDC system includes, but is not limited to, equipment operating parameters, process parameters, measurement data, and / or environmental and wafer information.

[0057] Machine operating parameters include, but are not limited to, data on mechanical status, electrical parameters, and environmental indicators. Mechanical status includes data such as vibration spectrum, bearing wear data, and robotic arm displacement accuracy; electrical parameters include voltage fluctuation values ​​(converted to percentage deviation), current load, and electrode voltage / current; environmental parameters include reaction chamber vacuum, gas flow rate (e.g., etching machine), and laser power (lithography machine). Process parameters include, but are not limited to, data on thin film processes, etching processes, and lithography processes. Thin film processes include data such as thickness distribution and refractive index in chemical vapor deposition; etching processes include data such as etching gas flow rate / pressure and plasma density / electron temperature; lithography processes include data such as exposure energy, focusing accuracy, and wafer stage vibration frequency (lithography machine). Measurement data includes, but is not limited to, electrical testing, optical testing, and surface morphology data. Electrical testing data includes chip current / voltage parameters; optical testing includes wavelength / intensity (optoelectronic devices); and surface morphology data includes device surface roughness and flatness (CMP process). Environmental and wafer information includes, but is not limited to, environmental data and wafer identification data. Environmental data includes temperature, humidity, and cleanliness, while wafer identification data includes serial number, size, and material batch.

[0058] In summary, the FDC system can collect relevant data from multiple dimensions through various configured sensors. For example, it collects mechanical status data through vibration and displacement sensors (monitoring robotic arm accuracy, bearing wear, etc.), electrical parameters through voltage / current sensors (measuring electrode voltage, current load, etc.), environmental indicators through vacuum sensors, gas flow meters, and laser power meters, thin film process data through thickness sensors and refractive index meters, and etching process data through plasma density sensors and electronic thermometers. The FDC system uses this collected data for fault and defective product detection, obtaining data related to machine malfunctions and product defects.

[0059] In current technologies, within the Forms Distributed Control (FDC) system of semiconductor manufacturing processes, FDC statistics / feature data are representative quantitative indicators extracted from raw time-series data. These indicators are used to reduce the dimensionality of high-dimensional, lengthy, and complex raw data into low-dimensional, information-dense feature data, enabling subsequent models to perform efficient processing and pattern recognition. For example, Unique Cell Difference Analysis (UVA) data is a common type of FDC statistics / feature data. As an indicator data processed using advanced computation, it is primarily used to quantify and monitor process consistency between different "processing units" within the same process equipment, thereby promptly identifying potential equipment failures or process anomalies. However, this type of processed data faces significant limitations in application. Taking FDC UVA data as an example, its limitations are mainly reflected in the following three aspects:

[0060] (1) The current process of generating FDC UVA data mainly relies on engineers' practical experience to configure parameters and select algorithms. Although this human experience-based approach can identify obvious process deviations to a certain extent, its standardization and reproducibility are insufficient. Different engineers may use different processing standards, resulting in a lack of uniformity in data results and affecting the reliability of subsequent analysis.

[0061] (2) In the process of processing raw data to generate FDC UVA data, some data features will inevitably be lost. This loss of features may lead to the omission of some important indicative process features, thereby affecting the timeliness and accuracy of anomaly detection.

[0062] (3) Due to the possible loss of information during the processing, the anomaly detection based on FDC UVA data cannot fully cover all potential fault modes, thus creating a detection blind spot.

[0063] The method proposed in this embodiment uses raw data for anomaly detection, which has significant technical advantages compared to relying solely on processed FDC statistics / feature data. By directly analyzing raw sensor data, this embodiment not only avoids information loss during data processing but also fully utilizes the complete feature information in the underlying data. This method of obtaining data directly from the source ensures the integrity of the analysis process and the reliability of the results, providing a more reliable technical guarantee for the stability of semiconductor manufacturing processes and product yield.

[0064] S200. Construct a historical data trace for each sensor based on the historical FDC data (a data trace is defined as a line reflecting the data trajectory obtained by connecting or fitting the data sequence recorded by a sensor during a complete process). For example, in a 30-second etching process, the temperature sensor records one value per second, and connecting these 30 data points constitutes a data trace for that temperature sensor.

[0065] In this step, the data will also be aligned, and the non-vibration curves formed by the data collected by the same sensor at different times in the historical data trace will be aligned according to the processing steps.

[0066] Data traces acquired by the same sensor at different times may experience a global shift due to systematic errors. This shift manifests as two data traces that should have similar shapes moving uniformly and consistently along the time axis, rather than exhibiting random, localized fluctuations. In semiconductor manufacturing, data traces recorded by the same sensor at different times may not completely overlap, but rather exhibit a global "shift." By aligning the different curves at each processing step, the interference caused by this shift can be effectively reduced, improving the accuracy and reliability of subsequent anomaly detection algorithms.

[0067] For example, in sensor data monitoring, if a specific characteristic point (such as peak temperature) in the temperature curve shows a time difference in different process operations (e.g., at the 10th second and the 12th second), this is considered a time axis offset. Such offsets are mainly caused by differences in the response delay of the process control system, which may involve the opening and closing lag of the temperature control valve or dynamic changes in the PID control parameters.

[0068] Without data alignment, directly comparing two data traces that contain shifts will result in systematic misjudgments. For example, the following situations may occur:

[0069] (1) Missing potential problems: A data trace has an abnormal spike in a critical process step, but the spike is "diluted" by the overall offset, making it impossible to detect effectively.

[0070] (2) Inaccurate clustering / modeling: For example, the clustering algorithm may classify two curves with the same shape but with offset as completely different categories, thus failing to correctly construct the first baseline library.

[0071] When performing alignment operations, the entire data trace is not treated as a whole. Instead, each data trace is divided into multiple independent data segments according to the process recipe's steps, and alignment is performed on each independent data segment of each process step. For example, a mean / median alignment method can be used to adjust the mean or median of all traces in the same step to the same level. Alternatively, a Dynamic Time Warping (DTW) algorithm can be used to simultaneously handle the scaling and offset of the numerical and time axes, finding the optimal matching path between two curves.

[0072] Given that each step in a semiconductor process has independent physical and chemical objectives (such as etching rate, deposition thickness, etc.), this embodiment innovatively employs a step-level alignment strategy: based on the recipe step definition, the entire data trace is decomposed into multiple independent process segments, and alignment operations are performed in a smaller, more stable local context. Compared to directly processing the entire complex and variable global curve, this method has the following significant advantages:

[0073] (1) Improved accuracy: By decomposing the complex global alignment problem into multiple simple local alignment sub-problems, non-critical systemic biases can be effectively eliminated, making real process anomalies stand out.

[0074] (2) Mechanism adaptability: Each process segment corresponds to a specific physicochemical process (such as pre-cleaning, main etching, and over-etching), and its data characteristics have inherent consistency. Step-level alignment can adopt an appropriate alignment algorithm for the characteristics of different process stages (such as mean normalization is suitable for steady-state processes, and DTW is suitable for dynamic change stages).

[0075] (3) Baseline library construction optimization: By eliminating systematic offsets between steps, a more accurate baseline model can be established, providing a reliable reference benchmark for anomaly detection.

[0076] (4) Engineering implementation advantages: The computational cost of local alignment is much smaller than that of global alignment, which can greatly reduce computational complexity and enhance robustness. It can also support incremental updates, allowing for the separate updating of the baseline model of a specific process segment without reprocessing the entire curve.

[0077] S300. Construct anomaly judgment rules for data trace. The judgment result of data trace includes normal data and abnormal data.

[0078] In this embodiment, the data trace is classified according to its shape characteristics, which can be further subdivided into vibration curves and non-vibration curves (i.e., data traces other than vibration curves, such as: regular curves, straight lines, etc.). Common anomalies in non-vibration curves typically include abnormal level values, abnormal jump points, and abnormal shifts. Please refer to [link to relevant documentation]. Figure 2 (a) Figure 2 (b) and Figure 2(c) shows an example of anomalies in the level value, jump point, and shift values ​​of a non-vibration curve. Anomalies in vibration curves mainly include frequency anomalies and amplitude anomalies. Please refer to [link to relevant documentation]. Figure 3 (a) and Figure 3 (b) shows an example of frequency anomalies and amplitude anomalies in the vibration curve, respectively.

[0079] Accordingly, the anomaly determination rules include a first anomaly determination rule corresponding to the non-vibration curve and a second anomaly determination rule corresponding to the vibration curve; the anomaly determination rules of these data traces will be illustrated exemplarily below.

[0080] For example, the first anomaly determination rule based on non-vibration curves such as conventional curves and straight lines can be configured as follows:

[0081] Calculate the similarity between the non-vibration curve to be judged and each class within the corresponding dataset in the pre-constructed first baseline library (i.e., the average similarity between the non-vibration curve to be judged and each non-vibration curve in that class). If the similarity between the non-vibration curve to be judged and each class is less than the corresponding similarity threshold, then mark the data of the non-vibration curve to be judged as abnormal; otherwise (i.e., the similarity between the non-vibration curve to be judged and at least one class is greater than or equal to the corresponding similarity threshold), mark the data of the non-vibration curve to be judged as normal.

[0082] The method for constructing the first baseline library is as follows:

[0083] Non-vibration curves in historical data traces of the Tool-Recipe-Sensor hierarchy are clustered using a clustering algorithm. Specifically, data from the same sensor under the same recipe within the same Tool are treated as a single dataset. The non-vibration curves in the historical data traces of each dataset are then clustered using the clustering algorithm, resulting in one or more clusters within each dataset. After removing clusters identified as noise (i.e., Cluster 0 obtained by the algorithm), a first baseline library is obtained, distributed as datasets containing these clusters.

[0084] In the Tool-Recipe-Sensor hierarchy, "Tool" refers to a specific production equipment, such as an etching machine. "Recipe" refers to a specific processing procedure or formula that runs on that machine, specifying how to process a product. "Sensor" refers to the various sensors installed on the machine. This hierarchical relationship means that the data analyzed is the historical data record of all sensors on a specific machine running a specific formula.

[0085] In this embodiment, the clustering algorithm preferably uses the DBSCAN algorithm. DBSCAN is a very popular density-based clustering algorithm. Its main advantage is that it does not require prior knowledge of the number of historical "normal" patterns, i.e., it does not require pre-specifying the number of clusters. Furthermore, the DBSCAN algorithm considers a cluster to consist of a group of densely connected points, while points not belonging to any dense region are marked as noise (usually Cluster 0), thus effectively identifying noise points. These noise points may be: genuine equipment failures or process anomalies, transient spikes captured by sensors, or certain rare, unclassifiable production conditions. When using the DBSCAN algorithm to cluster all historical data traces, the DBSCAN algorithm automatically discovers natural groupings within these complex data trajectories. Traces with similar sensor reading change patterns are grouped into the same cluster.

[0086] When using the DBSCAN algorithm, all data that cannot be categorized into any meaningful dense region (i.e., noise) is typically marked as Cluster 0. In other words, data in Cluster 0 is anomalous and unrepresentative production data (potentially stemming from brief equipment malfunctions, sensor false alarms, non-standard operations, etc.). Removing this noisy data ensures the "purity" of the baseline library, preventing the use of anomalous data as the standard for normal production.

[0087] The baseline library is a reference database composed of "healthy" and "normal" production data trajectories. It serves as the comparison standard for subsequent anomaly detection and fault diagnosis. By retaining all other non-noise clusters such as Cluster 1, Cluster 2, and Cluster 3, these clusters represent several major, stable production states that have historically occurred with this tool-recipe combination. This baseline library is no longer a single "standard curve" but is distributed in multiple groups. This makes the baseline library more reflective of the actual production situation (having multimodal characteristics). When new production data comes in, it can be determined which group's baseline it is most similar to, or whether it deviates from all groups, thus enabling more accurate fault detection.

[0088] Therefore, this embodiment uses an unsupervised learning algorithm (DBSCAN) to automatically learn normal process modes from historical data and improves the quality of the benchmark library by removing noise. This results in a more realistic and robust multimodal benchmark library, which can lay a solid foundation for subsequent real-time monitoring and anomaly detection, and greatly improve the intelligence level and accuracy of the FDC system.

[0089] The similarity threshold can be calculated using the following method:

[0090] Calculate the similarity (e.g., cosine similarity) of non-vibration curves within each class in each dataset of the first baseline library, and use the mean of the pairwise similarity between non-vibration curves within each class as the similarity threshold for that class.

[0091] Due to the special nature of vibration curves, they are not suitable for determination using the above methods. Therefore, this embodiment also constructs a second anomaly determination rule corresponding to the vibration curve. The second anomaly determination rule may include the following steps:

[0092] S311. Calculate the frequency of the vibration curve to be judged, and use an anomaly detection algorithm to calculate whether the frequency of the vibration curve is normal relative to the vibration curves in the second baseline library. The second baseline library is formed by analyzing the vibration curves in the historical data traces of each sensor.

[0093] The formula for calculating the frequency of the vibration curve is as follows:

[0094]

[0095] in, tf Indicates frequency; number or periods This indicates the number of periods in the vibration curve; Time This indicates the duration corresponding to the vibration curve.

[0096] The preferred anomaly detection algorithm used in this step is the Local Outlier Factor (LOF) algorithm. However, other anomaly detection algorithms can be selected as needed. LOF is a classic and powerful unsupervised anomaly detection algorithm. Like DBSCAN, LOF is a density-related algorithm, but its core idea is more refined: LOF does not compare the absolute density of a point, but rather compares the local density of a point with the local density of its neighbors to identify anomalies. This makes it particularly suitable for datasets with uneven density. Therefore, LOF is more suitable than DBSCAN for anomaly identification in vibration curves. Furthermore, LOF retains the advantages of unsupervised learning, requiring no prior knowledge of anomalous patterns.

[0097] S312, please refer to Figure 4 A sliding time window is introduced to calculate the local amplitude of the vibration curve to be judged within each time window (i.e., the difference between the maximum and minimum values ​​of FDC data within the time window). An anomaly detection algorithm is used to calculate whether the local amplitude of the vibration curve within each time window is normal relative to the vibration curve in the second baseline library.

[0098] Assuming the time window length is set to W (W can be an appropriate value as needed; if W is too small, the window vector can only capture very short-term trends, which may not be able to detect meaningful pattern anomalies and is more sensitive to noise; while if W is too large, it may "dilute" short-lived anomalies, making local density calculations inaccurate and increasing computational cost), starting from the beginning of the vibration curve, the window slides forward one time step each time, and each slide extracts a segment of length W from the vibration curve. When the duration of the vibration curve is T and the time step size is 1, a (T-W+1) segment of the curve will be extracted.

[0099] The preferred anomaly detection algorithm used in this step is the Looping-of-Flight (LOF) algorithm; however, other anomaly detection algorithms can be selected as needed. The standard LOF algorithm is designed for static, independently and identically distributed data points. It assumes that each data point is an independent vector, with no order or dependency between points. However, the data trace in this application is time-series data, characterized by strong temporal correlation between data points, making order crucial. The value of a point lies not only in itself but also in the local morphology or trend formed by its relationship with points before and after it. By introducing a sliding time window, one-dimensional time-series points can be transformed into a series of high-dimensional vectors representing "local morphology," thus preparing the input data for the LOF algorithm.

[0100] A second baseline library can be used to establish baselines for vibration curves. By collecting a large number of vibration curves from normal production, and processing them using a sliding window and LOF algorithm, the LOF value of each window vector can be calculated, revealing the distribution of LOF values ​​under normal conditions. By combining the sliding time window and the LOF algorithm, contextual anomalies can be captured, detecting both point anomalies and pattern anomalies. For example, for point anomalies, when a single point deviates significantly from normal values ​​(such as sensor spikes), a sufficiently large time window can capture it. For pattern anomalies / collective anomalies (i.e., a single point appears normal, but the pattern formed by a series of points is abnormal; for example, a normally stable signal suddenly oscillates at high frequencies), single-point detection methods are ineffective, while the combination of the sliding window and the LOF algorithm is best suited to detect this type of anomaly.

[0101] During online monitoring, a sliding window is used to extract a vibration curve to be judged. In real-time calculations, the new window vector is compared with the window vectors in the second baseline library to quickly calculate the LOF value of the new window. If the LOF value of a new window is significantly higher than the normal baseline (e.g., exceeding the 99th percentile of the baseline LOF value), it indicates a data anomaly, and it can be clearly identified at which stage of formulation execution the anomaly occurred.

[0102] However, the above methods have high computational complexity. Due to the large overlap of time windows, the number of generated vectors is huge, and LOF itself has high computational complexity, which may lead to performance bottlenecks. Therefore, when using the above methods, it is generally necessary to reduce the sampling frequency.

[0103] In this embodiment, a sliding time window is introduced into the LOF algorithm, which essentially elevates the anomaly detection problem in time series from a "point-to-point" comparison to a "local pattern to local pattern" comparison. This allows LOF to leverage its powerful relative density comparison capability to discover those "oddly shaped" segments in the time series, greatly enhancing its ability to detect complex anomaly patterns, making it very suitable for monitoring vibration curves.

[0104] S313. When the frequency and local amplitude of the vibration curve to be judged are both normal, the data of the vibration curve to be judged is marked as normal; otherwise, the data of the vibration curve to be judged is marked as abnormal, thereby completing the judgment on whether there is data abnormality in the vibration curve.

[0105] S400. During the production process, acquire real-time data from various sensors on the machine tool; for example, real-time data from various sensors on the machine tool can be acquired from the FDC system. Based on the real-time acquired data, construct a real-time data trace for each sensor as a data trace to be determined. The real-time data trace is obtained from the data corresponding to the currently completed process; for example, for a 30-second etching process, the sensor data during those 30 seconds will be used to form a real-time data trace only when the etching process is completed.

[0106] S500. Based on the constructed anomaly determination rules, determine whether the real-time data trace of each sensor is abnormal, obtain the first determination result for each sensor, and archive it. In this step, for non-vibration curves in the real-time data trace, the first anomaly determination rule is used for determination; for vibration curves in the real-time data trace, the second anomaly determination rule is used for determination, and the determination results are archived to provide a basis for subsequent sensor classification.

[0107] S600: When the archived first judgment results reach a predetermined quantity (e.g., data from three months of machine operation), a correlation analysis can be performed on the first judgment results and related quality results to determine the correlation between the first judgment results of each sensor and the corresponding product quality data. That is, when the sensor's first judgment result is normal, is the final processed product's quality data considered good? When the first judgment result is abnormal, is the final processed product's quality data considered defective? Thus, the correlation between the sensor's detection results and the machine's operating status is confirmed by whether the sensor's detection results correspond to the product quality. Based on this correlation, each sensor is classified to obtain its corresponding level.

[0108] In this step, a decision model can be used to determine the correlation and classify the sensors. In this embodiment, the decision model is preferably a random forest model. The random forest algorithm can systematically analyze the correlation between sensor FDC data anomalies and product quality, uncovering key sensors that may affect product quality beyond traditional engineer experience, thus supporting subsequent graded control of all sensors. This allows for timely warnings of related quality data after key sensor FDC data anomalies occur.

[0109] Please see Figure 5 The principle of the Random Forest algorithm is as follows: It constructs a large number of decision trees and trains them by randomly selecting different subsets of data through autonomous sampling. Finally, it uses voting or averaging to perform classification or regression tasks. Specifically, at each root node, the decision tree selects one feature from all features (random forests randomly select a subset) and uses this feature to split the dataset into two subsets, maximizing the separation between classes. This process is recursively repeated on each subset until a stopping condition is met, such as reaching the maximum depth, all data points belonging to the same class, or no more features available for splitting.

[0110] In this embodiment, the training dataset uses the FDC anomaly detection results (i.e., normal or abnormal) of each sensor under the same machine, process formula (product), and their corresponding quality data (Good or Bad). When applying the algorithm, a decision tree is built with the root node representing whether the sensor's FDC data is abnormal and the leaf nodes representing the probability of the quality data being Bad, thus constructing a random forest model. Once the model training is complete and the accuracy is high, the importance of each feature (whether the sensor is abnormal) is calculated as the degree of correlation (i.e., the strength of the correlation) between the sensor's FDC data and the quality data.

[0111] In this embodiment, when classifying sensors based on correlation, the sensors are divided into three levels. Level 1 consists of sensors whose first judgment result is strongly correlated with product quality, classified as high-risk sensors. Level 2 consists of sensors whose first judgment result is weakly correlated with product quality, classified as low-risk sensors. Level 3 consists of sensors whose first judgment result is unrelated to product quality, classified as no-risk sensors.

[0112] Based on the above embodiments, the machine sensor classification method of the present invention has the following unexpected technical effects:

[0113] (i) By systematically analyzing the correlation between abnormal detection results of machine FDC data and related quality data, this invention can more accurately find key sensors that may affect product quality, and the sensor classification is more accurate, providing a theoretical basis for subsequent machine management.

[0114] (ii) Classification is based on the sensor’s original data FDC Trace (i.e., sensor data trace) in the FDC data, which will not lose key information in the original data and the basis for classification is more comprehensive.

[0115] (iii) The grading results are largely unrelated to the engineers’ human experience, greatly reducing the reliance on human intervention in the grading process and avoiding the impact of engineers’ lack of experience on the grading.

[0116] Please see Figure 6 , Figure 6 A flowchart of a machine management method based on sensor data is shown. In the illustrated embodiment, the machine management method based on sensor data includes the following steps:

[0117] S910. The machine sensor classification method of any of the above embodiments is used to classify each sensor of the machine to obtain the corresponding level of each sensor. That is, the machine sensors are divided into three levels: high-risk sensors, low-risk sensors, and no-risk sensors.

[0118] S920. During the production process, acquire real-time data from each sensor on the machine, and construct a real-time data trace for each sensor based on the acquired data as the data trace to be judged. The specific process of this step is the same as step S400 in the sensor classification method.

[0119] S930. Based on the constructed anomaly determination rules, determine the second determination result corresponding to the real-time data Trace. The anomaly determination rules in this step are the data Trace anomaly determination rules constructed in step S300 in the machine sensor classification method of any of the above embodiments. The method for obtaining the second determination result in this step is the same as the method for obtaining the first determination result in step S500 of the sensor classification method; the difference is that the first determination result is only archived as the basis for subsequent sensor classification and is not used for machine control, while the second determination result is mainly used for machine control.

[0120] Of course, in this step, when the second judgment result indicates that the data is normal, the real-time data trace can be added to the baseline library (i.e., non-vibration curves are added to the first baseline library, and vibration curves are added to the second baseline library), thereby updating the baseline library to automatically adapt to slow changes caused by aging of machine parts, sensors, and other devices. During this process, older historical data traces in the baseline library can also be deleted to avoid an excessive amount of data in the baseline library. Conversely, if the second judgment result indicates that the data is abnormal, it can also be stored in the abnormal detection result archive.

[0121] S940. Implement machine control based on the corresponding level of each sensor and the second judgment result. In this step, because high-risk sensors are highly correlated with quality data, any anomaly is highly likely to affect product quality. Therefore, when the corresponding level of a sensor is a high-risk sensor and the second judgment result shows N anomalies, the machine is considered to be in an abnormal operating state, and abnormal control is implemented. N is generally a small positive integer, for example, N can be 1. That is, detecting even one anomaly in a high-risk sensor will determine that the machine's operating state is abnormal, thereby triggering abnormal control (e.g., stopping the machine and conducting an inspection). Because low-risk sensors have a lower correlation with quality data, it is impossible to determine whether a single anomaly will affect product quality. Therefore, when the corresponding level of a sensor is a low-risk sensor and the second judgment result shows M consecutive anomalies, the machine is considered to be in an abnormal operating state, and abnormal control is implemented. M is a positive integer greater than N. That is, only when multiple consecutive anomalies of a low-risk sensor are detected will the machine's operating state be determined to be abnormal, and abnormal control will be triggered. Since there is little correlation between risk-free sensors and quality data, the second judgment result of risk-free sensors is checked and evaluated periodically and is not used as the basis for judging the real-time working status of the machine. That is, after detecting an anomaly of a risk-free sensor, the machine's working status will not be judged to be abnormal immediately to avoid false alarms. However, the impact of the detected anomaly on quality will be checked and evaluated periodically. The machine's working status will be judged based on the periodic evaluation results, and control measures will be implemented if necessary.

[0122] To adapt to special scenarios and avoid misjudgments, when the second judgment result is a data trace anomaly, the following steps can also be performed:

[0123] S950. Determine whether PM (i.e., preventive maintenance) was performed before the data trace anomaly occurred. If so, it indicates that the anomaly may be a false alarm. In this case, using the current time point as the cutoff point, obtain the FDC data of relevant sensors before and after multiple (e.g., 5) preventive maintenance time points prior to this point.

[0124] Maintenance (PM) is a scheduled maintenance process for production equipment, involving planned maintenance, parts replacement, and sensor calibration. PM alters the equipment's "state." For example, after replacing a gas flow meter with a new one, even with the same setpoint, the actual flow curve may differ slightly from the old one. Therefore, the baseline readings of sensors often shift or fluctuate after PM. Since the "normal range" set by the flow meter control (FDC) system is based on historical data prior to PM, after PM, the data from relevant sensors may shift to a new, stable level. This new level may exceed the monitoring limits of the pre-PM baseline, triggering an "abnormal" alarm. In this case, it's not simply a matter of assuming a machine problem; rather, a judgment needs to be made based on the fluctuations in the relevant sensor FDC data along with the machine's PM readings.

[0125] S960. Calculate the fluctuation range of the FDC data of the relevant sensors before and after each preventive maintenance time node based on the FDC data, and calculate the average fluctuation range of the FDC data of the relevant sensors before and after a single preventive maintenance. Adjust the anomaly judgment rules of the data trace of the relevant sensors based on the average fluctuation range. For example, by analyzing the FDC data of the relevant sensors before and after the previous 5 machine PMs, a new "baseline" can be defined for the data after this PM based on the fluctuation after each PM, thereby avoiding false alarms.

[0126] Of course, in order to continuously optimize the anomaly detection rules of the sensor, the following steps can also be performed:

[0127] S970. Check if there are defective products when no abnormal machine operation is detected. If there are defective products, it means that there are still abnormal states of the machine that have not been detected. At this time, the sensors can be reclassified by random forest decision model to improve detection accuracy.

[0128] Based on the above embodiments, the machine control method based on sensor data of the present invention has the following unexpected technical effects:

[0129] (i) This embodiment differs from traditional control methods by using the sensor's original data FDC Trace (i.e., sensor data trace) in the FDC data for anomaly detection. Compared with using FDC statistics / feature data for detection, it does not lose key information in the original data, has a more comprehensive detection range, and more accurate detection results.

[0130] (ii) The control method is largely independent of the engineer’s human experience, greatly reducing the reliance on human intervention in the control process and avoiding the need for manual statistical control of all sensors, thus reducing the burden on human resources.

[0131] (iii) By systematically analyzing the FDC Trace data anomaly detection results with quality data such as Inline, Defect, WAT, and CP, key sensors associated with product quality data can be identified, thereby providing timely warnings of relevant quality data.

[0132] (iv) Based on the different degrees of correlation between sensor FDC data and product quality data, different risk levels are set for all sensors on the machine. Appropriate control rules are established based on historical data and abnormal situations. High-risk sensors are strictly controlled, and non-risk sensors are regularly evaluated. This can complement the current SPC system and greatly improve the quality control effect.

[0133] (v) Different detection methods are used for data traces in different scenarios (i.e., ordinary scenarios and PM scenarios) and different forms (i.e., non-vibration curves and vibration curves) to make the detection methods more suitable for the corresponding scenarios or curve forms, more targeted, and able to improve the accuracy of anomaly detection, enabling more effective and precise control, and greatly reducing false alarms and missed alarms.

Claims

1. A method for classifying machine sensor types, characterized in that, Includes the following steps: Acquire historical FDC data of the machine; Construct a historical data trace based on the data from each sensor in the historical FDC data; Construct rules for anomaly detection in data traces; Acquire real-time data from each sensor on the machine, and construct a real-time data trace for each sensor based on the acquired data; Based on the constructed anomaly determination rules, determine whether the real-time data trace of each sensor is abnormal, obtain the first determination result of each sensor, and archive it; When the number of archived first judgment results reaches a predetermined number, the correlation between the first judgment result of each sensor and the corresponding product quality data is determined, and each sensor is classified based on the correlation to obtain the corresponding level of each sensor. The data trace includes non-vibration curves, and the anomaly determination rule includes a first anomaly determination rule corresponding to the non-vibration curves; the first anomaly determination rule is configured as follows: The similarity between the non-vibration curve to be judged and each class within the corresponding dataset in the pre-constructed first baseline library is calculated. If the similarity with each class is less than the corresponding similarity threshold, the non-vibration curve to be judged is marked as abnormal; otherwise, the non-vibration curve to be judged is marked as normal.

2. The machine sensor classification method according to claim 1, characterized in that, The method for constructing the first baseline library includes the following steps: Align the non-vibration curves formed by data collected by the same sensor at different times in the historical data trace according to the processing steps; Using data from the same sensor under the same recipe of the same tool as a dataset, non-vibration curves in the historical data traces of each dataset are clustered based on a clustering algorithm, thereby obtaining one or more classes in each dataset; after removing classes identified as noise, a first baseline library is obtained in the form of datasets containing classes. The similarity threshold is calculated using the following method: Calculate the similarity of non-vibration curves within each class in each dataset of the first baseline library, and use the mean of the pairwise similarity between non-vibration curves within each class as the similarity threshold for that class.

3. The machine sensor classification method according to claim 1, characterized in that, The data trace includes a vibration curve, and the anomaly determination rule includes a second anomaly determination rule corresponding to the vibration curve. The second anomaly determination rule is configured as follows: The frequency of the vibration curve to be judged is calculated, and an anomaly detection algorithm is used to determine whether the frequency of the vibration curve is normal relative to the vibration curves in the second baseline library; wherein, the second baseline library is formed by vibration curves in the historical data trace of each sensor. A sliding time window is introduced to calculate the local amplitude of the vibration curve to be judged within each time window. An anomaly detection algorithm is used to calculate whether the local amplitude of the vibration curve within each time window is normal relative to the vibration curve in the second baseline library. If the frequency and local amplitude of the vibration curve to be judged are both normal within each time window, then the data of the vibration curve to be judged is marked as normal; otherwise, the data of the vibration curve to be judged is marked as abnormal.

4. The machine sensor classification method according to claim 1, characterized in that: In the step of determining the correlation between the first determination result of each sensor and the product quality data, and classifying each sensor based on the correlation to obtain the corresponding level of each sensor, a random forest model is used to determine the correlation and classify the sensors.

5. The machine sensor classification method according to claim 1, characterized in that: In the step of classifying each sensor based on the correlation, sensors whose first determination result is strongly correlated with product quality are classified as high-risk sensors, sensors whose first determination result is weakly correlated with product quality are classified as low-risk sensors, and sensors whose first determination result is unrelated to product quality are classified as no-risk sensors.

6. A machine control method based on sensor data, characterized in that, Includes the following steps: The sensor classification method of the machine tool according to any one of claims 1 to 5 is used to classify each sensor of the machine tool to obtain the corresponding level of each sensor; Acquire real-time data from each sensor on the machine, and construct a real-time data trace for each sensor based on the acquired data; Based on the constructed anomaly determination rules, it is determined whether the real-time data trace of each sensor is abnormal, and a second determination result is obtained; The machine is controlled based on the corresponding level of each sensor and the second judgment result.

7. The machine control method based on sensor data as described in claim 6, characterized in that, The step of controlling the machine based on the corresponding level of each sensor and the second determination result includes: When the corresponding level of the sensor is a high-risk sensor and the second judgment result is N abnormalities, the machine is considered to be in an abnormal working state, and abnormal control is implemented on the machine; when the corresponding level of the sensor is a low-risk sensor and the second judgment result is M consecutive abnormalities, the machine is considered to be in an abnormal working state, and abnormal control is implemented on the machine; where M is greater than N, and N and M are both positive integers. When the corresponding level of the sensor is a risk-free sensor, the second determination result of the risk-free sensor is checked and evaluated periodically. Based on the periodic evaluation results, it is determined whether the working status of the machine is abnormal, and corresponding control measures are implemented.

8. The machine control method based on sensor data according to claim 7, characterized in that: The value of N is 1.

9. The machine control method based on sensor data according to claim 6, characterized in that, When the second determination result is a data trace anomaly, the following steps are also performed: Determine whether preventative maintenance was performed before the data trace anomaly occurred; if so, take the current time point as the cutoff point and obtain the FDC data of the relevant sensors before and after the multiple preventative maintenance time points prior to this point. Based on the FDC data, calculate the fluctuation range of the FDC data of the relevant sensor before and after each preventive maintenance time node, and calculate the average fluctuation range of the FDC data of the relevant sensor before and after a single preventive maintenance. The anomaly detection rules for data traces from relevant sensors are adjusted based on the average fluctuation amplitude.

Images