A semiconductor motion platform temperature and pressure data fault diagnosis method based on formula non-stationary fault injection enhancement

Abstract

The application discloses a semiconductor motion table temperature and pressure data fault diagnosis method based on formula non-stationary fault injection enhancement. The method first takes real normal operation temperature and pressure data as a baseline, generates labeled enhanced samples according to five types of non-stationary abnormal forms, namely, slow decay of gas source pressure, intermittent gas impact, servo time-varying oscillation, thermal gas coupling drift and composite non-stationary fault, by using a formulaized disturbance function; then extracts pressure statistical features, temperature slow variable features, pressure difference impact features, temperature and pressure correlation and lag coupling features based on a multi-time scale sliding window to form a multi-scale temperature and pressure fusion feature vector; finally, an integrated tree diagnostic instrument is used to complete anomaly detection and fault type identification, and outputs fault categories, confidence and key contribution features. The application can establish an implementable motion table state diagnosis process when real fault samples are insufficient, and improve the detectability and fault type distinguishing ability of semiconductor motion table temperature and pressure non-stationary abnormalities.

Description

Technical Field

[0001] This invention relates to the fields of semiconductor equipment condition monitoring, data enhancement, and intelligent fault diagnosis, and particularly to a semiconductor motion stage temperature and pressure data fault diagnosis method based on formulaic non-stationary fault injection enhancement. Background Technology

[0002] Semiconductor motion stages are key components in lithography, inspection, packaging, and precision positioning equipment. Their air-bearing pressure supply system, servo feedback system, and thermal stabilization system jointly determine motion positioning accuracy, repeatability, and long-term equipment reliability. In actual operation, air pressure attenuation, short-term air interruption, servo loop oscillation, thermal drift, and multi-source composite anomalies often manifest as non-stationary characteristics such as slow drift, intermittent impacts, time-varying oscillations, hysteresis response, and multi-source superposition, rather than simple fixed threshold exceedances.

[0003] The scarcity of real-world fault samples is a major challenge in fault diagnosis for this type of equipment. Semiconductor equipment is high-value and operates in a controlled environment, making it difficult to proactively generate faults and collect a large number of labeled samples in actual production. Even when historical anomaly records exist, they are often few in number, imbalanced in categories, and incomplete in terms of fault stages. Traditional thresholding methods, 3σ methods, EWMA, and CUSUM are simple to implement but struggle to distinguish between normal slow drift and early fault evolution. Unsupervised methods such as PCA-SPE, Isolation Forest, and One-Class SVM can establish anomaly boundaries, but typically only output anomaly scores, failing to identify the fault type.

[0004] Therefore, a diagnostic method is needed that can still be implemented when there are insufficient real fault samples. This method should be able to construct interpretable enhanced fault samples based on real normal operation data, according to the air-bearing, servo and thermal stability operation mechanism of semiconductor motion stage, and combine multi-scale temperature and pressure fusion features to achieve anomaly detection and fault type identification. Summary of the Invention

[0005] The purpose of this invention is to provide a fault diagnosis method for semiconductor motion stage temperature and pressure data based on formulaic non-stationary fault injection enhancement. This method addresses the temperature and pressure detection data generated during the operation of a semiconductor motion stage. Under conditions of insufficient real fault samples, it constructs a diagnostic solution based on a process of "real normal data—formulaic fault injection—multi-scale temperature and pressure fusion features—integrated tree diagnostic output." This method is applicable to semiconductor equipment operation data analysis scenarios that are sensitive to supply pressure stability and thermal stability.

[0006] This invention provides a semiconductor motion stage temperature and pressure data fault diagnosis method based on formula-based non-stationary fault injection enhancement, comprising the following steps:

[0007] S1. Acquire multiple temperature and pressure signals generated during the operation of the target semiconductor motion stage, and construct real normal temperature and pressure baseline data from the normal operating status data;

[0008] S2. Based on the real normal temperature and pressure baseline data, a labeled enhanced sample is constructed using a formulaic non-stationary fault injection mechanism. The formulaic non-stationary fault injection mechanism includes at least one of the following: slow attenuation of gas supply pressure, intermittent gas interruption impact, servo time-varying oscillation, thermal gas coupling drift, and compound non-stationary fault.

[0009] S3. Set at least two different intensity levels for the same fault mechanism to generate multi-intensity enhanced samples, and combine the normal baseline samples, the new normal slow drift samples and the five types of fault enhanced samples to form a diagnostic sample set; the new normal slow drift samples are generated by superimposing a low-amplitude slow drift term not exceeding the allowable range of normal operating conditions on the real normal temperature and pressure baseline, and are used to simulate the slow change of environment or operating conditions under fault-free conditions.

[0010] S4. To address the difference between the fast response of pressure signals and the slow response of temperature signals, the diagnostic sample set is divided into multi-time-scale sliding windows to obtain short-window, medium-window, and long-window sample segments.

[0011] S5. Extract pressure statistical features, pressure differential impact features, temperature statistical and gradient features, temperature channel difference features, temperature-pressure correlation features, and temperature-pressure hysteresis coupling features from each window sample segment to form a multi-scale temperature-pressure fusion feature vector. Among them, the pressure statistical features include pressure mean, standard deviation, peak-to-peak value, and slope; the pressure differential impact features include maximum drop and differential standard deviation; the temperature statistical and gradient features include temperature mean, standard deviation, and slope; the temperature channel difference features include the maximum difference in temperature channels; and the temperature-pressure correlation features include temperature-pressure correlation and temperature-pressure differential correlation.

[0012] S6. Train a supervised diagnostic model based on the multi-scale temperature and pressure fusion feature vector, and perform binary classification anomaly detection and multi-class fault type identification.

[0013] S7. Process the temperature and pressure data of the semiconductor motion stage to be diagnosed according to steps S4 to S6, and output whether it is abnormal, the fault category, the category confidence level, and the key contribution characteristics.

[0014] In this invention, in step S2, a relative time ratio is defined, and a smoothing enhancement envelope function is used to represent the evolution of the fault from weak to strong over time:

[0015]

[0016] in, For the current moment, This marks the moment when the fault begins to evolve. The moment when the fault reaches a stable strength. This refers to a relative time proportion, i.e., a normalized time variable. To smooth and enhance the envelope function; when < When the fault intensity is 0, ≤ ≤ The fault intensity increases smoothly over time. > The fault reached a stable strength.

[0017] In this invention, in step S2, the gas supply pressure slowly decreases according to the following formula:

[0018]

[0019]

[0020] in, For the first Normal road pressure signal For the first Normal temperature signal of the road and These are the pressure and temperature signals after injection, respectively. To supply voltage attenuation amplitude, This is the pressure channel coefficient. The coefficient is slightly responsive to temperature. For temperature lag time, To smooth and enhance the envelope function, To account for the enhanced envelope function after temperature hysteresis.

[0021] In this invention, in step S2, the intermittent gas outage impact fault is injected according to the following formula:

[0022]

[0023]

[0024] in, The number of air-stop impacts. , and The first The center moment, duration, and relative intensity of the secondary impact. The gas interruption impact function is formed by the superposition of multiple Gaussian pulses. This is the normalized gas outage impact function. For the impact amplitude, This is the pressure channel coefficient. To avoid dividing by zero, small constants The pressure injected after the intermittent gas outage impact fault. For the first Normal road pressure signal.

[0025] In this invention, in step S2, the servo time-varying oscillation fault is injected according to the following formula:

[0026]

[0027]

[0028] in, and These are the initial and final oscillation frequencies, respectively. The timescale is the duration of the oscillation. For the initial phase, Let be the phase function whose frequency varies with time. The amplitude of the oscillation. This represents the pressure channel oscillation coefficient. For channel phase difference, To smooth and enhance the envelope function, Pressure injected after servo time-varying oscillation fault. For the first Normal road pressure signal.

[0029] In this invention, in step S2, the thermal gas coupling drift fault is injected according to the following formula:

[0030]

[0031]

[0032] in, This represents the slow drift amplitude of the pressure. This represents the slow temperature drift amplitude. The coupling coefficient between pressure drop and temperature hysteresis response is denoted as . This is the pressure channel coefficient. This is the temperature channel coefficient. For temperature lag time, and These represent the pressure and temperature after injection of heat-gas coupling drift fault data. For the first Normal road pressure signal For the first Normal temperature signal of the road For lag The first normal pressure signal after that, For lag The first injection pressure signal after that.

[0033] In this invention, in step S2, the composite non-stationary fault is formed by the superposition of at least two mechanisms among pressure attenuation, intermittent impact, servo oscillation, and temperature-pressure coupling drift, and its pressure channel injection formula is:

[0034]

[0035] in, The pressure injected after the complex non-stationary fault, For the first Normal road pressure signal , and These represent attenuation, impact, and oscillation amplitudes, respectively. and These are the enhanced envelopes for the decay term and the oscillation term, respectively. For the normalized gas outage impact function, It is the phase function of the time-varying oscillation. , and These are the pressure channel coefficients for attenuation, impact, and oscillation components, respectively. For channel phase difference, , and The weights of each fault component in a composite fault.

[0036] In this invention, in step S4, the short window, medium window, and long window correspond to different response time scales of short-term pressure impact, servo intermediate-frequency oscillation, and slow temperature and pressure drift, respectively, and the multi-scale temperature and pressure fusion feature vector of each diagnostic sample is represented as follows:

[0037]

[0038] in, For the first Multi-scale temperature and pressure fusion feature vector of a diagnostic sample , and These represent the values ​​at short, medium, and long window scales, respectively. Temperature and pressure features extracted from the window endpoints.

[0039] In this invention, step S5, the temperature-pressure hysteresis coupling characteristic includes temperature-pressure correlation coefficients under different hysteresis amounts:

[0040]

[0041] in, Lag quantity The temperature-pressure hysteresis correlation coefficient under these conditions This represents the difference or mean-removed representation of the pressure sequence within the window. relative pressure lag The temperature series is then represented by difference or mean removal. The correlation coefficient calculation function is... This is the hysteresis of temperature relative to pressure.

[0042] In this invention, in step S6, the supervised diagnostic model is trained using a labeled training set constructed from real normal baseline samples, new normal slow-drift samples, and fault enhancement samples generated according to the formulaic non-stationary fault injection mechanism in step S2. Anomaly detection and fault type identification are then achieved based on the multi-scale temperature and pressure fusion feature vector. The diagnostic output is:

[0043]

[0044] in, Output the fault category. For category confidence, For parameters of the supervised diagnostic model A defined category discrimination function, To supervise the diagnostic model in parameters The output class posterior probability or confidence level. For the first Multi-scale temperature and pressure fusion feature vector of a diagnostic sample.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] Enhanced samples are constructed based on real normal temperature and pressure data and a formulaic fault injection function, which can still generate labeled training data when real fault samples are insufficient.

[0047] By writing reproducible mathematical injection models for gas supply pressure attenuation, intermittent gas interruption, servo time-varying oscillation, thermo-gas coupling drift, and complex non-stationary faults, the interpretability and feasibility of the sample construction process are enhanced.

[0048] By simultaneously expressing fast pressure variables, slow temperature variables, and temperature-pressure hysteresis coupling relationships through short, medium, and long multi-scale windows, it can cover different fault evolution forms such as short-term shocks, medium-frequency oscillations, and slow drift.

[0049] In addition to anomaly detection, the fault category and key contribution characteristics are further output to help maintenance personnel determine the source of the anomaly.

[0050] The integrated tree diagnostic tool can be replaced with equivalent supervised models such as random forest, XGBoost, and gradient boosting tree, making it easy to deploy in engineering projects. Attached Figure Description

[0051] Figure 1 A schematic diagram for injecting five types of formulaic nonstationary faults.

[0052] Figure 2 This is a comparison chart of the performance of binary anomaly detection.

[0053] Figure 3 This is a confusion matrix for multi-class fault identification.

[0054] Figure 4 This is an experimental diagram of multi-scale thermo-baric coupling characteristic ablation.

[0055] Figure 5 This is a plot showing the importance of the top 20 features. Detailed Implementation

[0056] The invention will be further explained below with reference to real normal temperature and pressure data, the formulaic fault injection process, and experimental results. It should be emphasized that the data collected in this embodiment is the normal operating temperature and pressure baseline, and the five types of fault samples are all generated by injection based on the real normal baseline using the formulas listed in this specification. This embodiment does not use real equipment fault records, nor does it describe the injection sample results as generalization performance of real fault scenarios.

[0057] This invention provides a semiconductor motion stage temperature and pressure data fault diagnosis method based on formulaic non-stationary fault injection enhancement, the specific process of which is as follows.

[0058] Step 1: Constructing realistic normal temperature and pressure baseline data

[0059] The true normal baseline serves both as a reference for the fault-free state and as the original carrier for fault injection enhancement. In a specific embodiment, the acquired multi-channel temperature and pressure signals under stable operating conditions of the semiconductor motion stage can be denoted as:

[0060]

[0061] in, to For temperature channels, and This is a pressure channel.

[0062] Step 2: Construct five types of formulaic nonstationary fault injection functions

[0063] To avoid degrading fault enhancement to ordinary random noise, this invention constructs five types of formulaic non-stationary fault injection functions based on the coupling mechanism of air-bearing pressure supply, servo feedback, and thermal stability of semiconductor motion stages.

[0064] Fault intensity is enhanced by smoothing envelope function Control, short-term impact is controlled by Gaussian impulse function This indicates that time-varying oscillations are caused by a phase function whose frequency changes with time. The five types of faults are: slow decrease in air supply pressure, intermittent air supply interruption impact, servo time-varying oscillation, thermal-gas coupling drift, and combined non-stationary faults.

[0065] Based on the real normal temperature and pressure baseline data, this invention constructs labeled enhanced samples using a formulaic non-stationary fault injection mechanism. The formulaic non-stationary fault injection mechanism includes at least one of the following: slow attenuation of gas supply pressure, intermittent gas interruption impact, servo time-varying oscillation, thermal gas coupling drift, and compound non-stationary faults.

[0066] Specifically, we define the normalized time and employ a smoothing enhancement envelope function. This represents the evolution of a fault from weak to strong over time:

[0067]

[0068] in, This marks the moment when the fault begins to evolve. The moment when the fault reaches a stable strength.

[0069] The slow decrease in gas supply pressure fault is addressed using the following formula:

[0070]

[0071]

[0072] in, For the first Normal road pressure signal For the first Normal temperature signal of the road and These are the pressure and temperature signals after injection, respectively. To supply voltage attenuation amplitude, This is the pressure channel coefficient. The coefficient is slightly responsive to temperature. This refers to the temperature lag time.

[0073] The intermittent gas outage impact fault is injected according to the following formula:

[0074]

[0075]

[0076] in, The number of air-stop impacts. , and The first The center moment, duration, and relative intensity of the secondary impact. For the impact amplitude, This is the pressure channel coefficient. To avoid dividing by zero for small constants;

[0077] The servo time-varying oscillation fault is injected according to the following formula:

[0078]

[0079]

[0080] in, and These are the initial and final oscillation frequencies, respectively. The amplitude of the oscillation. This represents the pressure channel oscillation coefficient. This represents the channel phase difference.

[0081] The thermal-gas coupling drift fault is injected according to the following formula:

[0082]

[0083]

[0084] in, This represents the slow drift amplitude of the pressure. This represents the slow temperature drift amplitude. The coupling coefficient between pressure drop and temperature hysteresis response is denoted as . and These are the pressure and temperature channel coefficients, respectively.

[0085] The composite non-stationary fault is formed by the superposition of at least two mechanisms among pressure attenuation, intermittent impact, servo oscillation, and thermo-pressure coupling drift, and its pressure channel injection formula is:

[0086]

[0087] in, , and These represent attenuation, impact, and oscillation amplitudes, respectively. and These are the enhanced envelopes for the decay term and the oscillation term, respectively. , and The weights of each fault component in a composite fault.

[0088] In specific embodiments, the fault injection function can also be replaced with an equivalent function according to the equipment structure and sampling conditions. For example, the pressure attenuation can be replaced with a linear, nonlinear or piecewise attenuation function; the intermittent gas interruption can be replaced with a pulse function, a step recovery function or a random impulse function; the servo oscillation can be replaced with a time-varying frequency sinusoidal disturbance, a wavelet disturbance or other oscillation model; and the thermal gas coupling can be set with different lag times and coupling coefficients.

[0089] Step 3: Set up multiple intensity levels to generate enhanced samples

[0090] At least two different intensity levels are set for the same fault mechanism to generate multi-intensity enhanced samples. The normal baseline samples, new normal slow drift samples, and the five types of fault enhanced samples together constitute the diagnostic sample set. The new normal slow drift samples are generated by superimposing a low-amplitude slow drift term within the allowable range of normal operating conditions onto the real normal temperature and pressure baseline, used to simulate slow changes in the environment or operating conditions under fault-free conditions. This design enables the diagnostic model to learn the continuous evolution process from early minor anomalies to obvious faults. In a specific embodiment, weak, medium, and strong intensity levels are set for each type of fault, and the intensity coefficient is denoted as... And write the fault amplitude parameter as .

[0091] Step 4: Perform multi-scale sliding window partitioning

[0092] Because pressure signals respond quickly to gas path anomalies, while temperature signals respond slowly to thermal stability anomalies, this invention sets short, medium, and long windows. These windows correspond to different response time scales for short-term pressure shocks, servo intermediate-frequency oscillations, and slow temperature-pressure drift, respectively. Furthermore, the multi-scale temperature-pressure fusion feature vector for each diagnostic sample is represented as follows:

[0093]

[0094] in, , and These represent the short, medium, and long window lengths, respectively. Indicated by Temperature and pressure features extracted from the window endpoints.

[0095] In one embodiment, short window 60 seconds, medium window 120 seconds, long window The time interval is 300 seconds, with a step size of 60 seconds. For each window endpoint... Construct multi-scale feature vectors:

[0096]

[0097] In specific embodiments, the sliding window length and step size can be adjusted according to the device sampling frequency, gas path response time, and thermal response time.

[0098] Step 5: Extract multi-scale temperature and pressure fusion features

[0099] The following features are extracted at each scale: pressure mean, standard deviation, peak-to-peak value, slope, maximum drop, differential standard deviation, and dual-pressure channel difference; temperature mean, standard deviation, slope, and maximum temperature channel difference; temperature-pressure correlation, temperature-pressure differential correlation, and temperature-pressure hysteresis correlation.

[0100] The temperature-pressure hysteresis coupling characteristic includes the temperature-pressure correlation coefficient under different hysteresis values:

[0101]

[0102] in, This is the average value of the pressure channel. This is the average value across the temperature channels. This is the hysteresis of temperature relative to pressure.

[0103] Step 6: Train the supervised diagnostic model

[0104] A labeled training set is formed by injecting augmented samples, and an ensemble tree diagnostic tool is trained to perform binary anomaly detection and multi-class fault identification. A preferred embodiment uses an Extra Trees classifier. The classifier output may include whether an anomaly exists, the fault category, the category confidence score, and key contributing features. The ensemble tree diagnostic tool can also be replaced with random forests, extreme random trees, gradient boosting trees, XGBoost, support vector machines, or other supervised classifiers.

[0105] Step 7: Output the anomaly detection and fault type identification results.

[0106] The following are specific examples.

[0107] 1. Data Sources and Baseline Information

[0108] This embodiment uses a set of normal operating temperature and pressure data from a grating ruler servo motion stage as the baseline data. After reading, the data contains 5812 records, with an average sampling interval of approximately 1.000 seconds, a time range from 16:55:54.896 on February 5, 2026 to 18:32:45.900 on February 5, 2026, and a total duration of approximately 96.85 minutes. The data includes four temperature channels T1, T2, T3, and T4, and two pressure channels P1 and P2.

[0109] 2. Five Types of Formulaic Fault Injection Examples

[0110] This embodiment generates five types of fault-enhanced samples on a real normal baseline and sets intensity coefficients. The reference amplitudes for pressure attenuation, intermittent gas outage impact, and servo oscillation are 36 Pa, 44 Pa, 17 Pa, 20 Pa (pressure reference amplitude), and 0.075 °C (temperature reference drift amplitude) in thermal-gas coupling drift. Fault enhancement samples are shown below. Figure 1 As shown in the figure, the normal pressure curve represents the original true normal baseline, the pressure curve after injection represents the enhanced sample after superimposing disturbances according to the corresponding fault mechanism formula, and the intensity curve on the right represents the change of fault injection intensity over time. As can be seen from the figure, different fault types exhibit different non-stationary evolution characteristics in the pressure signal: slow attenuation of gas supply pressure is mainly manifested as a gradual decrease in pressure as the fault intensity increases; intermittent gas outage impact is manifested as a local short-term pressure drop; servo time-varying oscillation is manifested as periodic fluctuations that increase over time; thermal-gas coupling drift is manifested as slow pressure drift; and composite non-stationary faults simultaneously include multiple abnormal forms such as slow attenuation, impact, and oscillation. This figure illustrates that the fault samples of this invention are not simply randomly added noise, but rather interpretable enhanced samples constructed based on the equipment operating mechanism and formulaic disturbance functions. The parameters mentioned are used for the reproduction experiment in this embodiment and do not constitute a limitation on the scope of protection of this invention.

[0111] 3. Experimental Design

[0112] This embodiment constructs seven states: normal baseline, new normal slow drift, slow attenuation of gas supply pressure, intermittent gas outage impact, servo time-varying oscillation, thermo-gas coupling drift, and composite non-stationary fault. In the two-class classification task, normal baseline and new normal slow drift are classified as normal, while the five types of injection faults are classified as abnormal. In the multi-class classification task, the above seven states are classified.

[0113] The window sizes were set to 60s, 120s, and 300s, with a step size of 60s. Fault-type windows only retained segments with a maximum injection intensity of at least 0.08 within the window to avoid incorrectly labeling normal segments where faults have not yet occurred as faults. The training and test sets were hierarchically partitioned at the window level using a fixed random seed 20260514, with 70% allocated to the training set and 30% to the test set. This partitioning was used to verify the separability of the formulaic fault enhancement samples and multi-scale temperature-pressure fusion features, and is not intended as a claim to generalize across devices, batches, or real-world fault scenarios.

[0114] 4. Results of binary anomaly detection

[0115] The experimental results for the binary classification task are shown in the table below.

[0116] Table 1. Performance comparison of different methods in binary anomaly detection task.

[0117]

[0118] As shown in Table 1, the method of this invention, namely "multi-scale temperature and pressure fusion + Extra Trees", achieves Accuracy 0.9772, Precision 0.9725, Recall 0.9965, F1-score 0.9843, MCC 0.9431, and AUC 0.9966 in binary anomaly detection. These results are higher than XGBoost supervised diagnosis and single-scale features + Random Forest, indicating that the formulaic fault injection enhancement based on real normal data and multi-scale temperature and pressure fusion features can improve the separability between abnormal and normal states. Figure 2 The performance comparison results of different methods in binary anomaly detection tasks are also presented.

[0119] 5. Multi-class fault identification results

[0120] The experimental results for the multi-class classification task are shown in the table below.

[0121] Table 2. Performance comparison results of different methods in multi-class fault identification tasks

[0122]

[0123] As shown in Table 2, the method of this invention achieves an accuracy of 0.9188, a macro-F1 score of 0.8704, and a weighted-F1 score of 0.9123 in the seven state recognition categories, outperforming multi-scale statistical features + Extra Trees, XGBoost supervised diagnosis, and single-scale statistical features + Random Forest. This result indicates that in multi-class fault recognition tasks, relying solely on statistics or single-scale windows is insufficient to adequately distinguish different non-stationary fault types; the combination of temperature-pressure coupling features and multi-scale features contributes to fault type recognition.

[0124] 6. Ablation experiments and the importance of characteristics

[0125] To verify the contribution of multi-scale thermo-baric coupling features, this embodiment sets up multi-scale statistical features, single-scale statistical features, and XGBoost supervised diagnosis as controls. Figure 4 shows the ablation experimental results of multi-scale thermo-baric coupling features. Figure 4 As can be seen, the complete multi-scale thermo-pressure fusion feature achieved the highest Macro-F1 score of 0.8704, which is higher than the control scheme using only multi-scale and single-scale statistical features. This result indicates that the performance improvement is not solely due to the classifier, but is related to the joint expression of short, medium, and long multi-scale windows, slow temperature variable features, pressure shock features, and thermo-pressure hysteresis coupling features. The ablation experiment further demonstrates that the multi-scale thermo-pressure fusion feature system proposed in this invention makes a practical contribution to fault type identification.

[0126] The importance of the Top-20 features of the method of this invention is shown in the table below.

[0127] Table 3. Top-20 Feature Importance Results of the Method of the Invention

[0128]

[0129] As shown in Table 3, the mean temperature, peak-to-peak pressure, and standard deviation of pressure difference all contribute significantly to the long-term characteristics, indicating that long-term slow drift and pressure fluctuation characteristics play an important role in distinguishing different non-stationary states. This result is consistent with the coupled operation mechanism of the semiconductor motion stage's air-bearing pressurization and thermal stabilization system.

[0130] The above embodiment strictly generates fault enhancement samples based on real normal temperature and pressure data and the formulaic fault injection mechanism disclosed in this specification. Since real-world fault records of gas outages, servo anomalies, thermal drift anomalies, or combined anomalies are not used, the experimental results are only used to illustrate the feasibility and separability of the fault injection enhancement, feature extraction, and diagnostic process described in this invention on a real normal baseline. Parameters under real-world fault scenarios can be further calibrated based on subsequent real fault samples. Real fault data can be used as an independent test set or calibration set to further correct injection amplitude, impact width, oscillation frequency, temperature hysteresis time, coupling coefficient, window size, and classifier hyperparameters. In engineering deployment, this invention can first use real normal data and formulaic injection samples to form an initial diagnostic model, and then incrementally update it as real-world fault samples accumulate.

Claims

1. A semiconductor motion stage temperature and pressure data fault diagnosis method based on formulaic non-stationary fault injection enhancement, characterized in that, Includes the following steps: S1. Acquire multiple temperature and pressure signals generated during the operation of the target semiconductor motion stage, and construct real normal temperature and pressure baseline data from the normal operating status data; S2. Based on the real normal temperature and pressure baseline data, a labeled enhanced sample is constructed using a formulaic non-stationary fault injection mechanism. The formulaic non-stationary fault injection mechanism includes at least one of the following: slow attenuation of gas supply pressure, intermittent gas interruption impact, servo time-varying oscillation, thermal gas coupling drift, and compound non-stationary fault. S3. Set at least two different intensity levels for the same fault mechanism to generate multi-intensity enhanced samples, and combine the normal baseline samples, the new normal slow drift samples and the five types of fault enhanced samples to form a diagnostic sample set; the new normal slow drift samples are generated by superimposing a low-amplitude slow drift term not exceeding the allowable range of normal operating conditions on the real normal temperature and pressure baseline, and are used to simulate the slow change of environment or operating conditions under fault-free conditions. S4. To address the difference between the fast response of pressure signals and the slow response of temperature signals, the diagnostic sample set is divided into multi-time-scale sliding windows to obtain short-window, medium-window, and long-window sample segments. S5. Extract pressure statistical features, pressure differential impact features, temperature statistical and gradient features, temperature channel difference features, temperature-pressure correlation features, and temperature-pressure hysteresis coupling features from each window sample segment to form a multi-scale temperature-pressure fusion feature vector. Among them, the pressure statistical features include pressure mean, standard deviation, peak-to-peak value, and slope; the pressure differential impact features include maximum drop and differential standard deviation; the temperature statistical and gradient features include temperature mean, standard deviation, and slope; the temperature channel difference features include the maximum difference in temperature channels; and the temperature-pressure correlation features include temperature-pressure correlation and temperature-pressure differential correlation. S6. Train a supervised diagnostic model based on the multi-scale temperature and pressure fusion feature vector, and perform binary classification anomaly detection and multi-class fault type identification. S7. Process the temperature and pressure data of the semiconductor motion stage to be diagnosed according to steps S4 to S6, and output whether it is abnormal, the fault category, the category confidence level, and the key contribution characteristics.

2. The method for diagnosing temperature and pressure data faults on a semiconductor motion stage according to claim 1, characterized in that, In step S2, a relative time ratio is defined, and a smoothing enhancement envelope function is used to represent the evolution of the fault from weak to strong over time: ； in, For the current moment, This marks the moment when the fault begins to evolve. The moment when the fault reaches a stable strength. This is a relative time proportion. To smooth and enhance the envelope function; when < When the fault intensity is 0, ≤ ≤ The fault intensity increases smoothly over time. > The fault reached a stable strength.

3. The method for diagnosing temperature and pressure data faults on a semiconductor motion stage according to claim 2, characterized in that, For slow pressure decay faults in the gas supply, inject according to the following formula: ； ； in, For the first Normal road pressure signal For the first Normal temperature signal of the road and These are the pressure and temperature signals after injection, respectively. To supply voltage attenuation amplitude, This is the pressure channel coefficient. The coefficient is slightly responsive to temperature. For temperature lag time, To smooth and enhance the envelope function, To account for the enhanced envelope function after temperature hysteresis.

4. The method for diagnosing temperature and pressure data faults on a semiconductor motion stage according to claim 2, characterized in that, Intermittent gas outage impact faults should be injected according to the following formula: ； ； in, The number of air-stop impacts. , and The first The center moment, duration, and relative intensity of the secondary impact. The gas interruption impact function is formed by the superposition of multiple Gaussian pulses. This is the normalized gas outage impact function. For the impact amplitude, This is the pressure channel coefficient. To avoid dividing by zero, small constants The pressure injected after the intermittent gas outage impact fault. For the first Normal road pressure signal.

5. The method for diagnosing temperature and pressure data faults on a semiconductor motion stage according to claim 2, characterized in that, Servo time-varying oscillation faults are injected according to the following formula: ； ； in, and These are the initial and final oscillation frequencies, respectively. The timescale is the duration of the oscillation. For the initial phase, Let be the phase function whose frequency varies with time. The amplitude of the oscillation. This represents the pressure channel oscillation coefficient. For channel phase difference, To smooth and enhance the envelope function, Pressure injected after servo time-varying oscillation fault. For the first Normal road pressure signal.

6. The method for diagnosing temperature and pressure data faults on a semiconductor motion stage according to claim 2, characterized in that, The thermal gas coupling drift fault is injected according to the following formula: ； ； in, This represents the slow drift amplitude of the pressure. This represents the slow temperature drift amplitude. The coupling coefficient between pressure drop and temperature hysteresis response is denoted as . This is the pressure channel coefficient. This is the temperature channel coefficient. For temperature lag time, and These represent the pressure and temperature after injection of heat-gas coupling drift fault data. For the first Normal road pressure signal For the first Normal temperature signal of the road For lag The first normal pressure signal after that, For lag The first injection pressure signal after that.

7. The method for diagnosing temperature and pressure data faults on a semiconductor motion stage according to claim 2, characterized in that, Composite nonstationary faults are formed by the superposition of at least two mechanisms among pressure attenuation, intermittent impact, servo oscillation, and thermo-pressure coupling drift. The pressure channel injection formula is as follows: ； in, The pressure injected after the complex non-stationary fault, For the first Normal road pressure signal , and These represent attenuation, impact, and oscillation amplitudes, respectively. and These are the enhanced envelopes for the decay term and the oscillation term, respectively. For the normalized gas outage impact function, It is the phase function of the time-varying oscillation. , and These are the pressure channel coefficients for attenuation, impact, and oscillation components, respectively. For channel phase difference, , and The weights of each fault component in a composite fault.

8. The method for diagnosing temperature and pressure data faults on a semiconductor motion stage according to claim 1, characterized in that, In step S4, the short window, medium window, and long window correspond to different response time scales of short-term pressure shock, servo intermediate-frequency oscillation, and slow temperature and pressure drift, respectively, and the multi-scale temperature and pressure fusion feature vector of each diagnostic sample is represented as follows: ； in, For the first Multi-scale temperature and pressure fusion feature vector of a diagnostic sample , and These represent the values ​​at short, medium, and long window scales, respectively. Temperature and pressure features extracted from the window endpoints.

9. The method for diagnosing temperature and pressure data faults on a semiconductor motion stage according to claim 1, characterized in that, In step S5, the temperature-pressure hysteresis coupling characteristic includes the temperature-pressure correlation coefficient under different hysteresis amounts: ； in, Lag quantity The temperature-pressure hysteresis correlation coefficient under these conditions This represents the difference or mean-removed representation of the pressure sequence within the window. relative pressure lag The temperature series is then represented by difference or mean removal. The correlation coefficient calculation function is... This is the hysteresis of temperature relative to pressure.

10. The method for diagnosing temperature and pressure data faults on a semiconductor motion stage according to claim 1, characterized in that, In step S6, the supervised diagnostic model is trained using a labeled training set constructed from real normal baseline samples, new normal slow-drift samples, and fault enhancement samples generated according to the formulaic non-stationary fault injection mechanism in step S2. Anomaly detection and fault type identification are then achieved based on the multi-scale temperature and pressure fusion feature vector. The diagnostic output is: ； in, Output the fault category. For category confidence, For parameters of the supervised diagnostic model A defined category discrimination function, To supervise the diagnostic model in parameters The output class posterior probability or confidence level. For the first Multi-scale temperature and pressure fusion feature vector of a diagnostic sample.

Images