A method and system for fault monitoring of digital precision instruments and equipment
By acquiring multidimensional operating parameters and using a causal decoupling discrimination model to separate environmental and intrinsic contribution components, combined with a nonlinear trend evaluation model, the problem of difficulty in distinguishing fault characteristics caused by multi-physics coupling is solved, enabling reliable identification and effective maintenance of early equipment degradation characteristics.
Patent Information
- Application Number
- CN202611219169.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to effectively distinguish between fault characteristics caused by multi-physics coupling and operating condition drift. Furthermore, early degradation features are easily overshadowed by strong signals during incremental learning, leading to a decrease in the reliability and interpretability of equipment fault monitoring.
By acquiring multidimensional operating parameters, using a causal decoupling discriminant model to separate environmental contribution components from intrinsic contribution components, and combining this with a nonlinear trend assessment model, the health risk index is dynamically adjusted to generate maintenance recommendation information.
It enables reliable identification of early degradation characteristics of equipment, avoids the contamination of degradation trend analysis by environmental and operational interference, and provides clear equipment maintenance guidelines.
Smart Images

Figure CN123113799A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of equipment fault monitoring technology, and in particular to a method and system for fault monitoring of digital precision instruments and equipment. Background Technology
[0002] Precision instruments are widely used in semiconductor manufacturing, aerospace, and other fields, and their operational reliability directly determines product quality. Taking a lithography machine stage as an example, the equipment needs to perform high-speed movements with nanometer-level precision; any performance degradation can lead to huge losses. Traditional periodic maintenance methods are no longer sufficient to meet these needs, making equipment health monitoring a hot research topic.
[0003] Current fault monitoring methods mainly fall into two categories. One category is based on signal processing and machine learning, extracting fault features through methods such as wavelet transform and empirical mode decomposition before classification and identification. The other category is based on deep learning, using models such as convolutional neural networks to automatically learn fault modes, achieving some success in rotating machinery diagnosis. Fault detection and classification systems have been initially applied in some production lines.
[0004] However, existing technologies still have significant limitations. During the operation of precision instruments, signals from multiple physical fields, such as mechanical vibration, thermal stress, and electromagnetic interference, highly overlap in the frequency domain, making it difficult for traditional filtering to effectively separate interference without losing subtle fault characteristics. Deep learning models rely on a purely data-driven paradigm, learning statistical correlations rather than physical laws, leading to decreased reliability and a lack of interpretability when operating conditions change. Static threshold alarms, widely used in industrial settings, have high false alarm rates; introducing incremental learning makes strong signal samples easily cover early, subtle degradation characteristics, causing the system to gradually become insensitive to chronic performance degradation. Summary of the Invention
[0005] To address these issues, the present invention provides a method and system for monitoring faults in digital precision instruments and equipment, which overcomes the problems in the prior art where fault characteristics and operating condition drift are difficult to distinguish due to multi-physical field coupling, the causes of abnormalities cannot be traced, and early degradation characteristics are easily covered by strong signals and forgotten during incremental learning.
[0006] To achieve the above objectives, the present invention provides a method for fault monitoring of digital precision instruments and equipment, comprising: Step S1: Obtain multi-dimensional operating parameters during the operation of the target device. The multi-dimensional operating parameters include environmental field parameters, operating status parameters, and regional feature parameters of the first region on the target device. Step S2: Divide the regional feature parameters in the multidimensional operating parameters into several feature dimensions, extract the physical feature values of each feature dimension, obtain the theoretical benchmark value that matches the current operating state, calculate the deviation between each physical feature value and its corresponding theoretical benchmark value, and fuse each deviation to obtain the health risk index of the first region. Step S3: When the health risk index exceeds the preset risk threshold, the physical feature value is input into the causal decoupling discrimination model. The causal decoupling discrimination model separates the environmental contribution component and the intrinsic contribution component of the physical feature value according to the environmental field parameters, and outputs the anomaly type determination result of the first region. Step S4: Based on the abnormality type determination result, determine whether the exceedance of the health risk index belongs to environmental coupling abnormality or operation-related abnormality; if it belongs to environmental or operation-related early warning signal, output the environmental or operation-related warning signal; if it does not belong to environmental coupling abnormality, determine it as intrinsic abnormality and proceed to step S5. Step S5: Extract the time series of intrinsic contribution components of the first region within a preset historical window, calculate the degradation trend index of the time series using a nonlinear trend evaluation model, and output the fault risk level based on the matching degree between the degradation trend index and the aging curve of the equipment's entire life cycle.
[0007] Further, in step S1, the environmental field parameters include at least one of temperature field parameters, humidity field parameters, and electromagnetic field intensity parameters; The operating status parameters include at least one of the target device's movement speed, acceleration, load torque, and operating mode identifier; The regional characteristic parameters of the first region include at least one of the vibration signal, acoustic emission signal, deformation displacement signal, temperature distribution signal, and current / voltage signal of the first region.
[0008] Furthermore, in step S2, the physical characteristic values include time-domain statistical characteristics, frequency-domain energy distribution characteristics, and time-frequency domain joint characteristics; The theoretical benchmark value is determined based on historical data of the target equipment in a healthy state that matches the current operating conditions.
[0009] Furthermore, in step S3, the causal decoupling discrimination model is a neural network model embedded with physical constraint equations, wherein the physical constraint equations include at least one of the heat conduction equation, the dynamic equilibrium equation, and the electromagnetic field equation. The causal decoupling discrimination model takes the environmental field parameters and the operating state parameters as inputs, the environmental contribution component in the physical feature value as the output constraint, and the residual after subtracting the environmental contribution component from the physical feature value as the intrinsic contribution component.
[0010] Further, in step S3, the anomaly type determination result includes at least one of the following categories: The first type of anomaly is caused by the instantaneous fluctuation of environmental field parameters, which exceeds the standard of the health risk index. The second type of anomaly is caused by a transient event during the normal operation of the target device, where the health risk index exceeds the standard. The third type of anomaly is caused by the degradation of the intrinsic performance of the target device due to the exceeding of the health risk index.
[0011] Furthermore, in step S3, when determining the exception type, the following is also included: Obtain the operation instruction log of the target device, and determine whether there is a corresponding transient operation event when the health risk index exceeds the standard based on the timestamp information in the operation instruction log; If it exists, the exception type will be classified as a second type of exception.
[0012] Furthermore, in step S5, the nonlinear trend evaluation model adopts an incremental learning architecture based on elastic weight consolidation, and the model applies synaptic protection constraints to the historical degradation feature parameters of the first region during the update process.
[0013] Further, in step S5, the degradation trend index includes at least one of the monotonic change slope of the intrinsic contribution component time series, the coefficient of variation of fluctuation amplitude, and the energy proportion of high-frequency components. The matching degree is the similarity between the degradation trend index and the characteristic parameters of the corresponding stage in the aging curve of the entire life cycle of the equipment.
[0014] Furthermore, it also includes step S6, which generates maintenance suggestion information of different levels according to the fault risk level, and pushes the maintenance suggestion information to the terminal device.
[0015] This invention also provides a system for a digital precision instrument fault monitoring method, comprising: The data acquisition module is used to acquire multi-dimensional operating parameters of the target device during its operation. The feature extraction module, which is connected to the data acquisition module, is used to extract physical feature values of several dimensions of the first region according to the multidimensional operating parameters, and calculate the health risk index of the first region based on the deviation between the physical feature values and the corresponding theoretical benchmark values. The causal decoupling module, which is connected to the feature extraction module, is used to separate the physical feature value into an environmental contribution component and an intrinsic contribution component when the health risk index exceeds a preset risk threshold, and output the anomaly type determination result of the first region. An anomaly detection module, which is connected to the causal decoupling module, is used to determine the category of the health risk index exceeding the standard based on the anomaly type determination result. If it belongs to environmental coupling anomaly or operational correlation anomaly, it outputs an early warning signal. If it belongs to intrinsic anomaly, it triggers the trend analysis module. The trend analysis module, which is connected to the anomaly detection module, is used to calculate the degradation trend index based on the time sequence of the intrinsic contribution component after receiving a trigger, and output the fault risk level based on the matching degree between the degradation trend index and the aging curve of the equipment's entire life cycle. The output module, which is connected to the anomaly detection module and the trend analysis module respectively, is used to generate and push maintenance suggestion information based on the early warning signal or the fault risk level.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention obtains a theoretical benchmark value that matches the current operating state and calculates the deviation, enabling the assessment of the health risk index to be dynamically adjusted according to the actual operating state of the equipment, thus avoiding false alarms and missed alarms caused by traditional fixed thresholds under different operating conditions.
[0017] 2. This invention separates physical feature values into environmental contribution components and intrinsic contribution components by embedding a causal decoupling discrimination model of physical constraint equations. This allows interference components originating from the external environment or operational events in abnormal signals to be explicitly separated, and only the intrinsic contribution components are used for subsequent degradation analysis, thus solving the problem of being unable to trace the source of abnormalities.
[0018] 3. This invention pre-determines and outputs separately environmental coupling anomalies and operational correlation anomalies, ensuring that only intrinsic anomalies are included in the degradation trend assessment. This avoids the contamination of degradation trend analysis by environmental and operational interference, and achieves reliable identification of early degradation characteristics of equipment.
[0019] 4. This invention applies synaptic protection constraints to historical degenerate feature parameters through an incremental learning architecture based on elastic weight consolidation, enabling the model to retain the memory of early degenerate features while continuously learning new samples, thus overcoming the problem that early degenerate features are easily overwritten and forgotten by strong signals in incremental learning.
[0020] 5. This invention generates maintenance suggestion information, including maintenance time window, maintenance location, and maintenance operation type, based on the fault risk level, providing clear execution guidance for equipment maintenance. Attached Figure Description
[0021] Figure 1 This is a flowchart of a digital precision instrument fault monitoring method according to an embodiment of the present invention; Figure 2This is a schematic diagram of the module connection of the digital precision instrument and equipment fault monitoring system according to an embodiment of the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.
[0023] This invention uses a lithography machine workpiece stage as an example to illustrate the fault monitoring method for digital precision instruments and equipment. The lithography machine workpiece stage is a dual-stage system in a stepper lithography machine, which carries the wafer and moves it back and forth at high speed between the measurement station and the exposure station, performing nanometer-level precision stepper scanning at the exposure station. It should be understood that the lithography machine workpiece stage is only one specific example of the target equipment. This invention is also applicable to other digital precision instruments and equipment such as CNC machine tool spindles, aero-engine rotors, and precision measuring instrument stages. Those skilled in the art can adjust the sensor deployment scheme and the method of extracting physical feature values according to the specific type of target equipment.
[0024] Please see Figure 1 As shown, the fault monitoring method for digital precision instruments and equipment according to an embodiment of the present invention includes: Step S1: Obtain multi-dimensional operating parameters during the operation of the target device. The multi-dimensional operating parameters include environmental field parameters, operating status parameters, and regional feature parameters of the first region on the target device. Step S2: Divide the regional feature parameters in the multidimensional operating parameters into several feature dimensions, extract the physical feature values of each feature dimension, obtain the theoretical benchmark value that matches the current operating state, calculate the deviation between each physical feature value and its corresponding theoretical benchmark value, and fuse each deviation to obtain the health risk index of the first region. Step S3: When the health risk index exceeds the preset risk threshold, the physical feature value is input into the causal decoupling discrimination model. The causal decoupling discrimination model separates the environmental contribution component and the intrinsic contribution component of the physical feature value according to the environmental field parameters, and outputs the anomaly type determination result of the first region. Step S4: Based on the abnormality type determination result, determine whether the exceedance of the health risk index belongs to environmental coupling abnormality or operation-related abnormality; if it belongs to environmental or operation-related early warning signal, output the environmental or operation-related warning signal; if it does not belong to environmental coupling abnormality, determine it as intrinsic abnormality and proceed to step S5. Step S5: Extract the time series of intrinsic contribution components of the first region within a preset historical window, calculate the degradation trend index of the time series using a nonlinear trend evaluation model, and output the fault risk level based on the matching degree between the degradation trend index and the aging curve of the equipment's entire life cycle. Step S6: Generate maintenance suggestion information of different levels according to the fault risk level, and push the maintenance suggestion information to the terminal device.
[0025] Specifically, in step S1, the environmental field parameters include at least one of temperature field parameters, humidity field parameters, and electromagnetic field intensity parameters; The operating status parameters include at least one of the target device's movement speed, acceleration, load torque, and operating mode identifier; The regional characteristic parameters of the first region include at least one of the vibration signal, acoustic emission signal, deformation displacement signal, temperature distribution signal, and current / voltage signal of the first region.
[0026] In this embodiment of the invention, the first region is the working surface of the air-bearing guide rail of the lithography machine workpiece stage. Temperature field parameters are collected by a fiber optic grating temperature sensor array deployed on the workpiece stage base and surrounding environment, with a measurement accuracy of no less than ±0.01℃ and a sampling frequency of no less than 10Hz; humidity field parameters are collected by a high-precision humidity sensor, with a measurement accuracy of no less than ±1%RH; electromagnetic field strength parameters are collected by a triaxial electromagnetic field sensor, with a measurement frequency band covering 10Hz to 100MHz. The operating status parameters are read in real time by the workpiece stage control system via the EtherCAT industrial Ethernet protocol, including the workpiece stage scanning speed (0–0.8m / s), step acceleration (0–10m / s²), wafer load torque (0–5N·m), and the current working mode identifier (one of measurement mode, exposure mode, exchange mode, or zeroing mode). The vibration signal in the first region is acquired using a piezoelectric accelerometer with a frequency response range of 0.1Hz to 10kHz, installed at the four corners of the workpiece stage base and near the air-bearing guide slider. The acoustic emission signal is acquired using an acoustic emission sensor with a frequency response range of 20kHz to 1MHz, installed on the stator surface of the linear motor on the workpiece stage. The deformation displacement signal is acquired using an eddy current displacement sensor with a resolution of at least 0.1μm, installed on the X and Y axes of the workpiece stage, used to measure the static displacement deviation and dynamic fluctuations of the workpiece stage relative to the base. The temperature distribution signal is provided by a fiber optic grating temperature sensor array, with at least 32 temperature measurement points evenly spaced (20mm) along the length of the workpiece stage guide rail, used to capture the spatial gradient and temporal evolution of the temperature field on the guide rail surface. The current / voltage signal is acquired using a Hall current sensor and a voltage transformer, installed on the three-phase power supply line of the linear motor driver, with a sampling frequency consistent with the vibration signal (10kHz). Each sensor continuously acquires real-time data at a preset sampling frequency, and the data is aggregated to the feature extraction module via a data acquisition card.
[0027] The original sampling frequency of each sensor varies depending on the device type. The highest sampling frequency (10kHz) configured in the data acquisition module is used as the time reference. The data from other low-frequency sensors are upsampled to this reference frequency through linear interpolation. For high-frequency sensor data, the sampled value closest to the reference time is taken as the effective value for that time, thus unifying the sampling times of all sensors to the same time reference. For data points introduced by interpolation, no additional weight is added in the subsequent feature extraction; they only meet the model input requirements for temporal alignment. After spatiotemporal alignment, the data from each channel is divided according to a preset time window (1 second in this embodiment, i.e., each window contains 10,000 sampling points at a sampling rate of 10kHz). The timestamp of the last moment within the window is used as the timestamp of that window, and the time-stamped standardized tensor format feature data stream is output to the feature extraction module.
[0028] In this embodiment of the invention, the regional feature parameters are divided into three feature dimensions: vibration energy dimension, temperature field distribution dimension, and deformation displacement dimension. The vibration energy dimension takes the vibration signal as input and extracts the multidimensional physical feature values of the vibration signal in the time domain, frequency domain, and time-frequency domain; the temperature field distribution dimension takes the temperature distribution signal as input and extracts the spatial distribution characteristics and temporal evolution characteristics of the temperature field; the deformation displacement dimension takes the deformation displacement signal as input and extracts the static deviation and dynamic fluctuation characteristics of the displacement.
[0029] Specifically, in step S2, the physical feature values include time-domain statistical features, frequency-domain energy distribution features, and time-frequency domain joint features. Time-domain statistical features include statistical parameters such as mean, variance, peak value, kurtosis, root mean square value, waveform factor, and peak factor. Frequency-domain energy distribution features are extracted after converting the time-domain signal to the frequency domain using a fast Fourier transform, including parameters such as the energy proportion of each frequency band, dominant frequency amplitude, spectral centroid frequency, and spectral root mean square frequency. Time-frequency domain joint features are extracted using a short-time Fourier transform or wavelet transform, used to capture transient impact features in non-stationary signals.
[0030] The theoretical benchmark values are determined based on historical data of the target equipment in a healthy state that matches the current operating conditions. Specifically, the theoretical benchmark values are obtained as follows: after the target equipment is installed, debugged, and fully verified to be in a healthy state, feature data of each characteristic dimension under different combinations of operating states are collected to construct a health state benchmark database. The benchmark database uses operating state parameters as indexes and employs multidimensional interpolation or clustering methods to generate a corresponding theoretical benchmark value vector for each set of operating state parameters. When the equipment is running online, the theoretical benchmark value for the current operating state is retrieved or interpolated from the benchmark database based on the currently read real-time operating state parameters.
[0031] The deviation between each physical characteristic value and its corresponding theoretical reference value is calculated. The deviation is a measure of the degree to which the physical characteristic value deviates from the theoretical reference value. For each physical characteristic value of each characteristic dimension, the difference between it and the corresponding theoretical reference value is divided by the standard deviation of the characteristic value under healthy conditions to obtain the normalized deviation, thereby eliminating the influence of differences in dimensions and amplitudes between different characteristic dimensions.
[0032] The health risk index for the first region is obtained by fusing the various deviations. The fusing method is as follows: first, the deviations of each physical feature value under the same feature dimension are weighted and fused to obtain the health risk index component of that feature dimension; then, the health risk index components of each feature dimension are weighted and fused to obtain the health risk index for the first region. Each weight is determined using the entropy weight method or the analytic hierarchy process. The health risk index is a dimensionless quantity; the larger its value, the greater the degree to which the equipment deviates from its healthy state.
[0033] In some embodiments of the present invention, the time-domain statistical features, frequency-domain energy distribution features, and time-frequency domain joint features are extracted in the following manner: For time-domain statistical characteristics, statistical analysis is performed on the vibration signal sequence within a time window to extract statistical quantities that characterize the signal amplitude distribution characteristics, including the mean reflecting the average amplitude level of the signal, the variance reflecting the degree of signal dispersion, the kurtosis reflecting the kurtosis of the signal amplitude distribution, the root mean square value reflecting the effective energy of the signal, and the peak factor reflecting the signal impact characteristics.
[0034] To determine the frequency domain energy distribution characteristics, a fast Fourier transform is performed on the time domain signal after windowing to obtain the amplitude spectrum of the signal in the frequency domain. Then, the proportion of energy in each pre-divided frequency band to the total energy of the signal, as well as the centroid frequency and root mean square frequency of the spectrum, are calculated to characterize the concentration location and dispersion range of the signal energy in the frequency domain.
[0035] For the joint time-frequency domain features, wavelet transform is used to decompose the time-domain signal into multiple scales, extracting the energy proportion and its rate of change over time at each decomposition scale, so as to simultaneously characterize the joint distribution characteristics of the signal in both time and frequency dimensions.
[0036] Specifically, in step S3, the preset risk threshold is determined as follows: The equipment is continuously operated in a healthy state, and at least 1000 sets of normal data are collected. This sample size is sufficient to statistically characterize the distribution of the health risk index. The 95th percentile of the health risk index distribution is taken as the initial risk threshold. This quantile is a commonly used statistical benchmark in industrial process control, ensuring sufficient sensitivity to early abnormal signs while preventing misjudgment of normal data. Statistically, in this embodiment, the 95th percentile of the health risk index of the lithography machine stage in a healthy state is 0.35, and the initial risk threshold is set to 0.35. Based on this, considering the batch wafer scrapping loss caused by lithography machine stage failure and the downtime detection cost caused by a single false alarm, the ratio is approximately 5 to 10 times. Accordingly, the threshold is lowered by approximately 10% to 15% to improve the sensitivity to capturing early abnormal signs, and the preset risk threshold is finally adjusted to 0.30. When the health risk index exceeds 0.30, causal decoupling analysis is triggered.
[0037] Specifically, the causal decoupling discrimination model is a neural network model embedded with physical constraint equations. In this embodiment, the causal decoupling discrimination model adopts a three-layer fully connected network structure, taking temperature, humidity, and electromagnetic field strength from the environmental field parameters and scanning speed, acceleration, and load torque from the operating state parameters as inputs, and the environmental contribution component from the physical feature values as the output. The physical constraint equations include at least one of the following: heat conduction equation, dynamic equilibrium equation, and electromagnetic field equation. The heat conduction equation describes the diffusion evolution of the temperature field in the spatial and temporal dimensions, constraining the environmental contribution component in the temperature field distribution dimension to satisfy the heat diffusion law; the dynamic equilibrium equation describes the balance relationship between force, mass, and acceleration, constraining the environmental contribution component in the vibration energy dimension to satisfy Newton's second law; the electromagnetic field equation describes the coupling relationship between the electric field and the magnetic field, constraining the environmental contribution component in the current / voltage signal to satisfy the electromagnetic induction law.
[0038] Physical constraint equations are embedded in the model in the form of differentiable modules. The internal operations of each layer follow the corresponding physical equations. Layers are connected through standard tensor operations, and gradients can propagate normally in the backpropagation between layers, thus supporting gradient-based end-to-end training.
[0039] The total loss function of the causal decoupling discriminant model is a weighted average of data loss and physical constraint loss. Data loss measures the error between the predicted and actual measured values of the environmental contribution component, while physical constraint loss measures the deviation between the model output and the calculated values from the physical equations. Model training is performed by minimizing the total loss function, ensuring that the environmental contribution component output by the model both fits the data and conforms to physical laws. The weight coefficient of the physical constraint loss is used to adjust the proportion of physical constraints in the total loss. In this embodiment, this coefficient is between 0.05 and 0.1, based on the following criteria: using physical consistency and prediction accuracy on the validation set as evaluation indicators, a grid search is performed within the range of 0.01 to 0.5. When the weight is below 0.05, the model output is insufficiently constrained by physical constraints, and the prediction results deviate from physical laws when the data distribution shifts; when the weight is greater than 0.1, physical constraints excessively dominate the training process, and the model cannot fully fit the effective change patterns in the actual data. This range allows the model to achieve a balance between physical consistency and data fitting accuracy.
[0040] The intrinsic contribution component is the residual after subtracting the environmental contribution component from the physical characteristic value, reflecting the dynamic response characteristics of the equipment itself after excluding the influence of environmental factors.
[0041] Specifically, in step S3, the anomaly type determination result includes at least one of the following categories: The first type of anomaly is caused by instantaneous fluctuations in environmental field parameters when the health risk index exceeds the standard. If the environmental contribution component of the causal decoupling discrimination model accounts for more than 70% of the total contribution, and the change in environmental field parameters within 5 seconds before and after the exceeding time exceeds the preset environmental fluctuation threshold, it is judged as a first-type anomaly. The 70% threshold is determined as follows: using environmental coupling anomaly events in historical labeled data as samples, the actual distribution of the proportion of environmental contribution components to the total contribution in this type of event is statistically analyzed. The median is 78%, the 25th quantile is 65%, and 70% of the engineering tuning value between 65% and 75% is taken as the judgment threshold. This not only covers the vast majority of environmental coupling anomaly events, but also leaves sufficient margin for model output fluctuations to avoid misjudgments under normal environmental fluctuations. The preset environmental fluctuation thresholds are determined as follows: the temperature change threshold of 0.5℃ corresponds to three standard deviations of the environmental temperature control specification of 22±0.3℃ in the lithography machine workshop. This range covers 99.7% of normal temperature control fluctuations. Exceeding this range is considered a significant environmental fluctuation. The humidity change threshold of 5%RH corresponds to the upper limit of the allowable deviation of the workshop humidity control specification. The electromagnetic field strength change threshold of 20dBμV / m corresponds to the normal interference limit specified in the electromagnetic compatibility design of the equipment. Exceeding this value indicates the presence of a strong external interference source.
[0042] The second type of anomaly involves a health risk index exceeding the standard caused by transient events during normal operation of the target equipment. These transient events include rapid deceleration and zeroing of the workpiece stage, laser interferometer zeroing calibration, dual workpiece stage exchange, and working mode switching. If the difference between the time of the health risk index exceeding the standard and the timestamp of the aforementioned transient events is less than ±200ms, and the proportion of the operational correlation contribution component output by the causal decoupling discrimination model exceeds 50% of the total contribution, it is classified as a second type of anomaly. The ±200ms judgment window is based on the following: the timestamp accuracy of the operation command recorded by the workpiece stage control system is 1ms; there is a synchronization error of approximately ±5ms between the time reference accuracy of sensor data acquisition and the control system clock; and the duration of the transient event itself is 100ms to 300ms. The ±200ms window is sufficient to cover the entire process of the transient operation from the issuance of the command to the completion of the mechanical response, as well as the clock synchronization error between the two systems. The 50% threshold is determined based on the following: taking operation-related abnormal events in historical labeled data as samples, the actual distribution of the proportion of operation-related contribution components in the total contribution of such events is statistically analyzed. The median is 62%. 50% is set as the judgment threshold to ensure that more than half of the contribution components in normal operation transient events are attributed to the operation itself, and to avoid misjudging the weak environmental fluctuations that accompany normal operation transients as environmental coupling anomalies.
[0043] The third type of anomaly is caused by the degradation of the intrinsic performance of the target equipment when the health risk index exceeds the standard. If it does not meet the conditions of either the first or second type of anomaly (i.e., the environmental contribution component accounts for less than 70% and there is no corresponding operational transient event), it is judged as a third type of anomaly.
[0044] Specifically, obtaining the operation command log of the target device is one of the preliminary steps in anomaly type determination. The operation command log is generated by the lithography machine's workpiece stage control system and records the issuance time, command type, and command parameters of each operation command. The comparison results of the operation command log, along with the environmental contribution component and intrinsic contribution component output by the causal decoupling discrimination model, are used as input conditions for anomaly type determination. If the difference between the time of exceeding the health risk index and the timestamp of a certain transient operation event is less than ±200ms, that time is marked as the operation-related time. If multiple transient operation events exist within 500ms before and after the operation-related time, the one with the closest timestamp is taken as the matching event. The 500ms matching window is determined based on the fact that the dual-workpiece stage exchange process of the lithography machine involves multiple sub-actions, and the time interval between each sub-action is typically 100ms to 300ms. The 500ms window is sufficient to accommodate the command timestamps of all sub-actions during the entire exchange process, ensuring that the correct matching event closest to the time of exceeding the standard is obtained. If there is a sudden change in environmental field parameters, i.e., a temperature change exceeding 0.5℃, a humidity change exceeding 5%RH, or a sudden increase in electromagnetic field strength exceeding 20dBμV / m within 1 second, it is classified as a Type I anomaly. If both the operation-related moment and the sudden change in environmental field parameters exist simultaneously, it is classified as a Type II anomaly according to the operation-related priority principle, because operation-related events are usually accompanied by electromagnetic interference and temperature fluctuations, and are primarily attributed to operational transients.
[0045] In some embodiments of the present invention, the causal decoupling discriminant model outputs a confidence score for each anomaly category when outputting the anomaly type determination result. The confidence score is calculated as follows: for the first type of anomaly, the confidence score is the proportion of the environmental contribution component to the total contribution; for the second type of anomaly, the confidence score is the proportion of the operational association contribution component to the total contribution; and for the third type of anomaly, the confidence score is the proportion of the intrinsic contribution component to the total contribution. When the confidence score is below 60%, the determination result is an uncertain anomaly, triggering a secondary confirmation process. The 60% confidence threshold is determined based on the calibration curve between the actual proportion of each type of anomaly in historical labeled data and the model output probability. 60% corresponds to the inflection point interval in the calibration curve where the predicted probability deviates the least from the actual occurrence frequency. This threshold can retain the effective anomaly detection rate to the greatest extent without significantly increasing the false positive rate. The secondary confirmation process caches the physical feature value at this moment and outputs the final determination result only after the determination results of the subsequent three consecutive time windows are consistent, in order to avoid false positives caused by instantaneous disturbances.
[0046] Specifically, in step S4, if the anomaly type determination result is either a Type I or Type II anomaly, then the exceedance is determined to be either an environmental coupling anomaly or an operational correlation anomaly. For Type I anomalies, the environmental contribution component accounts for more than 70% of the total contribution, indicating that the increase in the health risk index is mainly caused by the instantaneous fluctuation of environmental field parameters, and the intrinsic state of the equipment has not undergone substantial degradation. For Type II anomalies, the difference between the time of the health risk index exceeding the standard and the timestamp of the operational transient event is less than ±200ms, indicating that the exceedance is caused by transient impacts or electromagnetic pulses generated during normal equipment operation. Both types of anomalies are caused by interference factors outside the equipment itself and should not trigger further analysis of equipment failure; only an environmental or operational correlation warning signal needs to be output.
[0047] Environmental or operational-related early warning signals include the warning type, the timestamp of exceeding the limit, the health risk index value, and suggested actions. When the warning type corresponds to the first type of anomaly, this warning signal is an environmentally related warning signal, and the suggested action is to check environmental parameters and verify whether there are abnormal fluctuations in temperature, humidity, or electromagnetic field strength. When the warning type corresponds to the second type of anomaly, this warning signal is an operationally related warning signal, and the suggested action is to check the equipment operation log to confirm whether there were normal operational transient events such as sudden deceleration to zero, zero calibration, dual workpiece stage exchange, or mode switching when the health risk index exceeded the limit. After this warning signal is pushed to the central control room display terminal and the maintenance personnel's mobile terminal, the process terminates and does not proceed to the subsequent intrinsic anomaly trend analysis steps.
[0048] If the anomaly type determination result is a Class III anomaly, meaning the exceedance of the health risk index is caused by intrinsic performance degradation of the target equipment, it is determined to be an intrinsic anomaly, and proceeds to step S5. Under this path, only the time series of the intrinsic contribution component is transmitted to the trend analysis module; the environmental contribution component and the operation-related contribution component do not participate in the subsequent trend analysis, thereby avoiding contamination of the intrinsic degradation trend assessment by external interference factors.
[0049] In this embodiment of the invention, considering that the performance degradation of the lithography machine stage usually occurs slowly on a weekly or monthly basis, the preset historical window in step S5 is set to 30 days. That is, with the current time as the endpoint, all time-series data of the intrinsic contribution components in the past 30 days are extracted, downsampled by the minute average, and then used for subsequent trend calculation.
[0050] Specifically, in step S5, the nonlinear trend evaluation model adopts an incremental learning architecture based on elastic weight consolidation. Its core principle is that when the model continuously learns new samples, it applies synaptic protection constraints to parameters with higher importance among the historical degradation feature parameters. This limits the update magnitude of these parameters by the training gradient of new samples to a preset update threshold, thereby protecting the model's memory of early degradation features from being overwritten by new samples. In this embodiment, importance is measured by the diagonal elements of the Fisher information matrix; a larger value indicates that the parameter is more important to the old task. The preset update threshold is the maximum allowable change in the importance weights of historical degradation feature parameters, set at 3%, meaning that the relative change of any important parameter in a single update does not exceed 3%. This value is based on the following: using historical operating data of similar equipment as the test set, parameter scanning is performed within the range of 1% to 10%. When the threshold is below 3%, the model cannot fully adapt to the normal fluctuations of the new operating conditions; when the threshold is above 3%, the retention accuracy of early degradation features begins to decline significantly. 3% is a balance point that takes into account both the adaptability to new operating conditions and the retention accuracy of old features.
[0051] The degradation trend indicators include at least one of the monotonic slope of the intrinsic contribution component time series, the coefficient of variation of fluctuation amplitude, and the energy proportion of high-frequency components. In this embodiment, all three indicators are calculated simultaneously.
[0052] The monotonic slope is calculated using the Theil-Sen estimator: the slope is calculated for all point pairs in the time series, and the median of all slopes is taken as the monotonic slope. This estimator is insensitive to outliers and is suitable for robust trend estimation of industrial data. A positive monotonic slope indicates continuous deterioration of equipment performance, while a negative slope indicates performance improvement. The larger the absolute value of the slope, the faster the degradation. This indicator is the core representation of intrinsic degradation trend.
[0053] The coefficient of variation (COP) of fluctuation amplitude is the ratio of the standard deviation to the mean, reflecting the stability of equipment operation within a historical window. A larger COP indicates more unstable equipment operation and more severe fluctuations. This indicator is used to capture changes in equipment operational stability; a sustained increase in the COP often foreshadows changes in the equipment's dynamic characteristics.
[0054] The proportion of high-frequency component energy is obtained by calculating the ratio of high-frequency sub-band energy to total energy after wavelet packet decomposition of the time series. An abnormal increase in the proportion of high-frequency component energy often indicates early signs of failure. This indicator is sensitive to weak high-frequency impact signals generated during the early degradation of equipment and can detect signs of degradation before the monotonic change slope shows a significant upward trend.
[0055] The matching degree is the similarity between the degradation trend index and the characteristic parameters of the corresponding stage in the equipment's full life cycle aging curve. The equipment's full life cycle aging curve is obtained through statistical analysis of historical operating data of similar equipment, describing the typical evolution trajectory of each degradation trend index throughout the entire process from equipment commissioning to scrapping. In this embodiment, the full life cycle aging curve of the lithography machine's workpiece stage is constructed based on historical data from multiple similar devices operating continuously for five to eight years, with the number of operating days as the horizontal axis and the typical value and its variation range of the degradation trend index as the vertical axis. The matching degree is measured using the reciprocal of the dynamic time warping distance, and the value range of the matching degree is 0 to 1. The closer the matching degree is to 1, the more closely the current degradation trend matches the typical aging trajectory; the lower the matching degree, the greater the degree to which the current degradation trend deviates from the typical aging trajectory.
[0056] Specifically, when the matching degree is less than the first preset matching degree threshold, it is judged as a high-risk level, and the output module generates emergency maintenance suggestions accordingly. When the matching degree is greater than or equal to the first preset matching degree threshold and less than the second preset matching degree threshold, it is determined to be of medium risk level, and the output module generates preventive maintenance suggestions. When the matching degree is greater than or equal to the second preset matching degree threshold, it is determined to be a low-risk level, and the output module maintains the normal monitoring state. The calibration method for the first preset matching degree threshold is as follows: based on the matching degree distribution of the degradation trend index and the full life cycle aging curve of the same type of equipment in the 30 days before the historical failure, the 10th quantile of this distribution is taken as the high-risk judgment threshold, that is, only 10% of the pre-failure data will be missed as non-high-risk. Based on the statistical results of historical failure cases of lithography machine workpiece stages, the average matching degree in the 30 days before the failure is about 0.55, and the 10th quantile is about 0.44. After on-site verification and adjustment, the first preset matching degree threshold is set to 0.50. This value takes into account the balance between the advance warning of high-risk warning and the false judgment rate: at this threshold, the average advance warning of high-risk warning is about seven days, while the false judgment rate is controlled within an acceptable range. The calibration method for the second preset matching degree threshold is as follows: the 90th quantile of the matching degree distribution under normal equipment operation is used as the boundary between medium risk and low risk. In this embodiment, it is statistically shown that the matching degree under normal operation is usually not less than 0.70, so the second preset matching degree threshold is set to 0.70.
[0057] Please see Figure 2 As shown, the system for the digital precision instrument and equipment fault monitoring method of this invention includes: The data acquisition module is used to acquire multi-dimensional operating parameters of the target device during its operation. The feature extraction module, which is connected to the data acquisition module, is used to extract physical feature values of several dimensions of the first region according to the multidimensional operating parameters, and calculate the health risk index of the first region based on the deviation between the physical feature values and the corresponding theoretical benchmark values. The causal decoupling module, which is connected to the feature extraction module, is used to separate the physical feature value into an environmental contribution component and an intrinsic contribution component when the health risk index exceeds a preset risk threshold, and output the anomaly type determination result of the first region. An anomaly detection module, which is connected to the causal decoupling module, is used to determine the category of the health risk index exceeding the standard based on the anomaly type determination result. If it belongs to environmental coupling anomaly or operational correlation anomaly, it outputs an early warning signal. If it belongs to intrinsic anomaly, it triggers the trend analysis module. The trend analysis module, which is connected to the anomaly detection module, is used to calculate the degradation trend index based on the time sequence of the intrinsic contribution component after receiving a trigger, and output the fault risk level based on the matching degree between the degradation trend index and the aging curve of the equipment's entire life cycle. The output module, which is connected to the anomaly detection module and the trend analysis module respectively, is used to generate and push maintenance suggestion information based on the early warning signal or the fault risk level.
[0058] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.
[0059] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A method of digitalized precision instrument equipment failure monitoring, characterized in that, include: Step S1: Obtain multi-dimensional operating parameters during the operation of the target device. The multi-dimensional operating parameters include environmental field parameters, operating status parameters, and regional feature parameters of the first region on the target device. Step S2: Divide the regional feature parameters in the multidimensional operating parameters into several feature dimensions, extract the physical feature values of each feature dimension, obtain the theoretical benchmark value that matches the current operating state, calculate the deviation between each physical feature value and its corresponding theoretical benchmark value, and fuse each deviation to obtain the health risk index of the first region. Step S3: When the health risk index exceeds the preset risk threshold, the physical feature value is input into the causal decoupling discrimination model. The causal decoupling discrimination model separates the environmental contribution component and the intrinsic contribution component of the physical feature value according to the environmental field parameters, and outputs the anomaly type determination result of the first region. Step S4: Based on the abnormality type determination result, determine whether the exceedance of the health risk index belongs to environmental coupling abnormality or operational correlation abnormality. If it belongs to the category, an environmental or operational warning signal will be output; if it does not belong to the category, it will be determined as an intrinsic anomaly and proceed to step S5. Step S5: Extract the time series of intrinsic contribution components of the first region within a preset historical window, calculate the degradation trend index of the time series using a nonlinear trend evaluation model, and output the fault risk level based on the matching degree between the degradation trend index and the aging curve of the equipment's entire life cycle.
2. The method of claim 1, wherein, In step S1, the environmental field parameters include at least one of temperature field parameters, humidity field parameters, and electromagnetic field intensity parameters; The operating status parameters include at least one of the target device's movement speed, acceleration, load torque, and operating mode identifier; The regional characteristic parameters of the first region include at least one of the vibration signal, acoustic emission signal, deformation displacement signal, temperature distribution signal, and current / voltage signal of the first region.
3. The method of claim 1, wherein, In step S2, the physical characteristic values include time-domain statistical characteristics, frequency-domain energy distribution characteristics, and time-frequency domain joint characteristics; The theoretical benchmark value is determined based on historical data of the target equipment in a healthy state that matches the current operating conditions.
4. The digital precision instrument equipment failure monitoring method of claim 1, wherein, In step S3, the causal decoupling discrimination model is a neural network model embedded with physical constraint equations, which include at least one of the heat conduction equation, dynamic equilibrium equation, and electromagnetic field equation. The causal decoupling discrimination model takes the environmental field parameters and the operating state parameters as inputs, the environmental contribution component in the physical feature value as the output constraint, and the residual after subtracting the environmental contribution component from the physical feature value as the intrinsic contribution component.
5. The method of claim 1, wherein, In step S3, the anomaly type determination result includes at least one of the following categories: The first type of anomaly is caused by the instantaneous fluctuation of environmental field parameters, which exceeds the standard of the health risk index. The second type of anomaly is caused by a transient event during the normal operation of the target device, where the health risk index exceeds the standard. The third type of anomaly is caused by the degradation of the intrinsic performance of the target device due to the exceeding of the health risk index.
6. The method of claim 5, wherein, In step S3, when determining the exception type, the following is also included: Obtain the operation instruction log of the target device, and determine whether there is a corresponding transient operation event when the health risk index exceeds the standard based on the timestamp information in the operation instruction log; If it exists, the exception type will be classified as a second type of exception.
7. The digitized precision-instrument device failure monitoring method of claim 1, wherein, In step S5, the nonlinear trend evaluation model adopts an incremental learning architecture based on elastic weight consolidation, and the model applies synaptic protection constraints to the historical degradation feature parameters of the first region during the update process.
8. The method of claim 1, wherein, In step S5, the degradation trend index includes at least one of the monotonic change slope of the intrinsic contribution component time series, the coefficient of variation of fluctuation amplitude, and the energy proportion of high-frequency components. The matching degree is the similarity between the degradation trend index and the characteristic parameters of the corresponding stage in the aging curve of the entire life cycle of the equipment.
9. The digitized precision-instrument device failure monitoring method of claim 1, wherein, It also includes step S6, which generates maintenance suggestion information of different levels according to the fault risk level, and pushes the maintenance suggestion information to the terminal device.
10. A system for performing the fault monitoring method for digital precision instruments and equipment according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire multi-dimensional operating parameters of the target device during its operation. The feature extraction module, which is connected to the data acquisition module, is used to extract physical feature values of several dimensions of the first region according to the multidimensional operating parameters, and calculate the health risk index of the first region based on the deviation between the physical feature values and the corresponding theoretical benchmark values. The causal decoupling module, which is connected to the feature extraction module, is used to separate the physical feature value into an environmental contribution component and an intrinsic contribution component when the health risk index exceeds a preset risk threshold, and output the anomaly type determination result of the first region. An anomaly detection module, which is connected to the causal decoupling module, is used to determine the category of the health risk index exceeding the standard based on the anomaly type determination result. If it belongs to environmental coupling anomaly or operational correlation anomaly, it outputs an early warning signal. If it belongs to intrinsic anomaly, it triggers the trend analysis module. The trend analysis module, which is connected to the anomaly detection module, is used to calculate the degradation trend index based on the time sequence of the intrinsic contribution component after receiving a trigger, and output the fault risk level based on the matching degree between the degradation trend index and the aging curve of the equipment's entire life cycle. The output module, which is connected to the anomaly detection module and the trend analysis module respectively, is used to generate and push maintenance suggestion information based on the early warning signal or the fault risk level.
