Component operation and maintenance prediction method, system and equipment of semiconductor equipment and medium

Abstract

The invention provides a component operation and maintenance prediction method and system for semiconductor equipment, equipment and a medium, and the method comprises the steps: collecting multi-factor operation data in an equipment operation process, carrying out the normalization processing of the multi-factor operation data, and forming a normalized task execution time length, normalized energy consumption data and a normalized current waveform sequence; based on the normalized multi-factor operation data, comparing corresponding data of the first operation arm and the second operation arm according to a task type and a residual time window to form a task duration residual feature, an energy consumption residual feature and a current waveform residual feature, and jointly constructing a mirror image residual feature vector; the mirror image residual feature vectors are input into a prediction model, the prediction model analyzes residual features of continuous time windows, an anomaly score set is generated, anomaly judgment is conducted on the working states of the fixed 180-degree opposed double operation arms through a double-threshold judgment mechanism, and anomaly result data are obtained; and outputting an early warning prompt based on the abnormal result data.

Description

Technical Field

[0001] This application relates to semiconductor technology, and more particularly to a method, system, device, and medium for predicting the maintenance of components in a semiconductor device. Background Technology

[0002] In semiconductor wafer manufacturing equipment, robotic arms are responsible for key actions such as gripping, handling, loading, and unloading wafers, and are core components for the normal operation of the production line. Currently, the most common robotic arm structures on the market include single-arm and dual-arm types. Among them, the fixed 180° opposing dual-arm design is widely used in the industry because of its relatively simple structure, lower cost, and higher efficiency compared to a single-arm design.

[0003] However, even fixed 180° opposed double manipulators can experience component wear, decreased energy efficiency, and unstable operation over long-term operation. Current technologies typically rely on single data points for assessment, such as whether a task execution time is excessively long or energy consumption exceeds limits. This approach is susceptible to occasional fluctuations or noise, resulting in either false alarms or missed alarms, failing to accurately reflect the true health status. More importantly, existing methods rarely utilize dynamic signals during operation, such as the waveform of current changes over time. These signals, which could reveal hidden problems like motor overload or jamming, are often overlooked.

[0004] Furthermore, traditional methods are mostly based on static threshold judgments and do not analyze the entire operation process of the operator arm in a time series context, making it difficult to detect potential degradation trends in the early stages. Traditional methods often wait until obvious equipment failures occur before the system issues an alarm, lacking proactive support for maintenance work.

[0005] Therefore, how to comprehensively utilize various operational information such as task execution time, energy consumption data, and current waveform sequences under the structural characteristics of fixed 180° opposing dual manipulators, and identify anomalies by comparing the differences between the two manipulators under the same task, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] This application provides a method, system, device, and medium for predicting the maintenance of components in semiconductor equipment, in order to solve the problems in the prior art.

[0007] In a first aspect, this application provides a method for predicting the maintenance of components in semiconductor equipment, including: Multi-factor operation data acquisition and preprocessing: During equipment operation, multi-factor operation data of fixed 180° opposing double operating arms are collected, including task execution time, energy consumption data and current waveform sequence. The multi-factor operation data is then normalized to form normalized task execution time, normalized energy consumption data and normalized current waveform sequence. The mirror residual feature vector is constructed by comparing the corresponding data of the first and second operating arms according to the normalized task execution time, normalized energy consumption data and normalized current waveform sequence, according to the task type and residual time window, to form task duration residual features, energy consumption residual features and current waveform residual features, and then constructing the mirror residual feature vector based on the three residual features. The prediction model is constructed and anomaly detection is performed by inputting the mirror residual feature vector into the prediction model. The prediction model analyzes the mirror residual feature vector of the continuous time window, generates anomaly score set, and uses a dual threshold judgment mechanism to judge the working state of the fixed 180° opposing double manipulators and obtain anomaly result data. Operation and maintenance prompts are generated, and early warning prompts are output based on the abnormal result data. The early warning prompts include operation and maintenance prompts and abnormal diagnosis prompts.

[0008] In one possible design, the construction of the mirror residual feature vector includes: Construct a residual time window, and within each residual time window, construct a task duration sample set, an energy consumption sample set, and a current waveform sample set based on the task type. Then, aggregate each sample set according to the operator arm identifier to form the first operator arm sample set and the second operator arm sample set. Finally, start residual feature analysis when the residual analysis conditions are met. In the task duration sample set, the median of the task duration samples of the first operating arm and the second operating arm and the absolute deviation of the median of the task duration samples are calculated to construct the task duration residual features, and standardized task duration residuals are formed through standardization processing. In the energy consumption sample set, the median of the energy consumption samples of the first operating arm and the second operating arm and the absolute deviation of the median of the energy consumption samples are calculated to construct the energy consumption residual characteristics, and standardized energy consumption residuals are formed through standardization processing. In the current waveform sample set, the median and absolute deviation of the current waveform sample of the first operating arm and the second operating arm are calculated to construct the current waveform residual features, and standardized current waveform residual features are formed through standardization processing. Under the same residual time window and task type, the standardized task duration residual, standardized energy consumption residual, and standardized current waveform residual are jointly constructed into a mirror residual feature vector.

[0009] In one possible design, the prediction model construction and anomaly detection include: A prediction model is constructed based on the time series characteristics of mirror residual feature vectors; The mirror residual feature vector, composed of standardized task execution time residual, standardized energy consumption residual, and standardized current waveform residual, is used as the input of the prediction model to predict and analyze the working state of the fixed 180° opposing dual operating arms, and outputs a multi-dimensional set of anomaly scores. Based on the aforementioned abnormal score set, a dual-threshold judgment mechanism combining deviation threshold and confidence threshold is adopted to determine the abnormal working state of the fixed 180° opposing dual manipulators.

[0010] In one possible design, the residual time window is a division of the continuous operation process of the manipulator into fixed time lengths; Within each residual time window, tasks falling into the residual time window are selected according to task type. The normalized task execution time, normalized energy consumption data and normalized current waveform sequence of each task are sampled as task duration sample, energy consumption sample and current waveform sample, respectively, to form a task duration sample set, energy consumption sample set and current waveform sample set. Based on the operator arm identification, the task duration sample set, energy consumption sample set, and current waveform sample set are collected to form the first operator arm sample set and the second operator arm sample set. By comparing the corresponding data of the first manipulator sample set with the second manipulator sample set, the residual characteristics of task duration, energy consumption, and current waveform are obtained.

[0011] In one possible design, the component maintenance prediction method is based on the structural characteristics of a fixed 180° opposing double operating arm. Under the same residual time window and task type, the normalized task execution time, normalized energy consumption data and normalized current waveform sequence are sampled to form corresponding task time samples, energy consumption samples and current waveform samples, which is the mirror operation. The task duration samples, energy consumption samples, and current waveform samples are collected according to the operator arm identifier to form the first operator arm sample set and the second operator arm sample set, and then compared. The difference between them is the mirror residual.

[0012] In one possible design, the dual threshold determination mechanism includes: When the comprehensive anomaly score is greater than or equal to the preset emergency threshold, the operating arm is determined to be abnormal and the anomaly level is marked as level three. When the overall anomaly score is less than the preset emergency threshold and greater than or equal to the preset deviation threshold, it is determined that there is an abnormal deviation in the current residual time window, and the continuous counter is incremented by 1. When the comprehensive anomaly score shows abnormal deviations in N consecutive residual time windows, the operating arm is determined to be abnormal and the anomaly level is marked as Level 1 anomaly. When the comprehensive anomaly score shows abnormal deviations in all M consecutive residual time windows, the operating arm is determined to be abnormal and the anomaly level is identified as Level 2 anomaly. When the overall anomaly score is less than the preset deviation threshold, the continuous counter is reset to zero. Where N and M are both positive integers, and M is greater than N, and the emergency threshold is greater than the deviation threshold.

[0013] In one possible design, the prediction model is a deep neural network model, optimized using parameter pruning and 8-bit integer quantization.

[0014] Secondly, this application provides a component maintenance prediction system for semiconductor equipment, comprising: The multi-factor operation data acquisition and preprocessing module acquires multi-factor operation data of the fixed 180° opposing double operating arms during equipment operation, including task execution time, energy consumption data and current waveform sequence, and performs normalization processing on the multi-factor operation data to form normalized task execution time, normalized energy consumption data and normalized current waveform sequence. The mirror residual feature vector construction module compares the corresponding data of the first and second operating arms according to the normalized task execution time, normalized energy consumption data and normalized current waveform sequence, according to the task type and residual time window, to form task duration residual features, energy consumption residual features and current waveform residual features, and jointly constructs a mirror residual feature vector based on the three residual features. The prediction model construction and anomaly detection module inputs the mirror residual feature vector into the prediction model. The prediction model analyzes the mirror residual feature vector of the continuous time window, generates anomaly score set, and uses a dual threshold judgment mechanism to judge the working state of the fixed 180° opposing double manipulators and obtain anomaly result data. The operation and maintenance prompt generation module outputs early warning prompts based on the abnormal result data. The early warning prompts include operation and maintenance prompts and abnormal diagnosis prompts.

[0015] Thirdly, this application provides an electronic device, comprising: Processor; and, Memory for storing the executable instructions of the processor; The processor is configured to perform any of the possible methods described in the first aspect by executing the executable instructions.

[0016] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement any of the possible methods described in the first aspect.

[0017] This application provides a visual monitoring method, system, device, and medium for semiconductor equipment. The method collects multi-factor operational data, including task execution time, energy consumption data, and current waveform sequences, during the operation of the semiconductor equipment. Under the structural condition of a fixed 180° opposing dual-operating arm, it constructs a mirror residual feature vector, reflecting the differences in task execution based on the symmetry comparison of the two operating arms. Furthermore, this method combines a predictive model with a dual-threshold judgment mechanism to achieve hierarchical anomaly detection of the operating arm's working status and generate maintenance and diagnostic prompts. This enables early identification of potential component degradation, reduces the risk of downtime due to failure, and improves the foresight and reliability of equipment maintenance.

[0018] Furthermore, by constructing a residual time window and grouping samples according to task type, the robustness of residual feature calculation is ensured; residual features are extracted using the median and absolute deviation of the median of task duration, energy consumption, and current waveform, enhancing anti-interference capability; a deep neural network is introduced into the prediction model, and parameter pruning and 8-bit integer quantization are used for lightweight optimization, ensuring the real-time inference capability of the prediction model on edge devices; at the same time, the dual threshold judgment mechanism effectively reduces the false alarm rate and false negative rate by setting deviation threshold and emergency threshold and continuous window counting method.

[0019] In summary, the visual monitoring method, system, device, and medium for semiconductor devices proposed in this application have the following beneficial effects: This application preprocesses task execution time, energy consumption data, and current waveform sequences by normalizing them, providing a unified basis for comparison in terms of speed, energy efficiency, and dynamics for different types of tasks. Based on the normalized data, a mirrored residual feature vector is constructed. In a fixed 180° opposed dual-operating-arm structure, comparative analysis of the differences between the two arms under the same task allows for earlier detection of signs of component performance changes. Combining a residual time window mechanism and a robust calculation method for median absolute deviation effectively reduces the interference of occasional fluctuations and noise on the results, making the residual features more stable and reliable. By introducing a lightweight deep neural network prediction model and running it in real-time at the edge, dynamic analysis of residual changes within continuous time windows enables timely identification of abnormal states and improves detection accuracy.

[0020] Therefore, this application not only improves the accuracy and real-time performance of anomaly detection, but also provides interpretable diagnostic information for maintenance personnel, significantly improving the overall maintenance efficiency and reliability of semiconductor equipment. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0022] Figure 1 This is a flowchart illustrating a component maintenance prediction method for semiconductor devices according to an example embodiment of this application; Figure 2 This is a schematic diagram illustrating the process of constructing mirror residual feature vectors according to an example embodiment of this application; Figure 3 This is a schematic diagram illustrating the process of predictive model construction and anomaly detection according to an example embodiment of this application; Figure 4 This is a schematic diagram of the structure of a component maintenance prediction system for semiconductor devices according to an example embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device according to an example embodiment of this application.

[0023] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0025] This embodiment proposes a component maintenance prediction method, system, device, and medium for semiconductor equipment, applicable to semiconductor equipment with fixed 180° opposed dual manipulators. In this scenario, the two manipulators typically work together in a mirror-image manner to complete wafer gripping, handling, loading, and unloading tasks. As operating time increases, the manipulator drive components, motors, and transmission mechanisms gradually deteriorate, manifesting as asymmetric differences in task execution time, energy consumption data, and current waveform sequences between the two manipulators. To address this, this application constructs mirror residual features and combines them with a lightweight prediction model to achieve early detection and maintenance alerts for component degradation.

[0026] Figure 1 This is a flowchart illustrating a component maintenance prediction method for semiconductor devices according to an example embodiment of this application. Figure 1 As shown, the component maintenance prediction method for semiconductor devices provided in this embodiment includes: In this embodiment, a "task" refers to an independent action unit issued by the semiconductor equipment control system and completed by the wafer transfer arm within one process cycle. Each task has a clear start and end timestamp, used to describe the complete execution process of the action. Typical tasks include actions such as gripping, handling, loading, and unloading.

[0027] To facilitate subsequent data processing, this embodiment introduces the following identifiers: Task ID: Represents the original task instance recorded in the semiconductor equipment control system log, used to uniquely identify the specific execution process of an operator arm within a process cycle; Arm ID: Applicable to fixed 180° opposing dual manipulators, used to distinguish the first manipulator (Arm_A) and the second manipulator (Arm_B). Task Type: Used to indicate the action category of a task, including grabbing, moving, loading and unloading. The task type is bound to the task identifier and written to the system log to clarify the action attributes of each task.

[0028] Step S101: Multi-factor operation data acquisition and preprocessing. During equipment operation, multi-factor operation data of the fixed 180° opposing dual operating arms are acquired, including task execution time, energy consumption data and current waveform sequence. The multi-factor operation data is then normalized to form normalized task execution time, normalized energy consumption data and normalized current waveform sequence.

[0029] In this step, during normal operation of the semiconductor device, multi-factor operational data of the fixed 180° opposed dual manipulators are collected. The fixed 180° opposed dual manipulators include a first manipulator (Arm_A) and a second manipulator (Arm_B). The multi-factor operational data includes task execution time, energy consumption data, and current waveform sequences. The specific acquisition process is as follows: Task execution time: Extract the start timestamp (t) of each task from the system log of the semiconductor device control system. start ) and end timestamp (t) end ), and calculate the execution time of each task:

[0030] in: ExecutionTime j The execution duration of task j; , is the start timestamp of task j; , is the end timestamp of task j; The execution time of each task is bound to the operator arm identifier (ArmID), task identifier (TaskID), and task type (TaskType).

[0031] Energy consumption data: The voltage and current of the first and second operating arms are collected by the power acquisition module of the semiconductor device or an external power sensor, and the data are collected at a fixed sampling period. Based on the voltage and current of the first and second manipulators, the instantaneous power of the first and second manipulators is calculated respectively, and then continuously spliced ​​to form the task execution interval [t]. start ,t end The instantaneous power curve; Within the task execution interval [t] start ,t end Integrating the instantaneous power curve within the given timeframe yields the total energy consumption data for this task. The calculation formula is as follows:

[0032] in: Total Energy j , where is the energy consumption data for task j; , is the start timestamp of task j; , is the end timestamp of task j; P(t) represents the instantaneous power of task j.

[0033] Each energy consumption data point (TotalEnergy) is bound to the operator arm identifier (ArmID), task identifier (TaskID), and task type (TaskType).

[0034] Current waveform sequence: By using a current detection module or Hall sensor in a semiconductor device, the current signals I(t) of the first and second manipulators are acquired in real time during task execution to obtain the task execution interval [t]. start ,t end The current waveform sequence, where: The current waveform sequence for TaskID=j is as follows:

[0035] in: CurrentWaveform j (t) represents the current waveform sequence of task j, where t is the sampling time during task execution. , is the sampling point for the current signal of the first operating arm; j, the task identifier is j; , is the start timestamp of task j; , is the end timestamp of task j; Each current waveform sequence is bound to a task identifier (TaskID) and a task type (TaskType).

[0036] The collected multi-factor operational data were preprocessed, including: Normalization process: For the execution time of each task, a min-max normalization method is used to generate a normalized task execution time, which is used to eliminate differences in the cycle time of different tasks.

[0037] Among them, ExecutionTime j Let ExecutionTime be the execution duration of task j. j ' is the normalized execution time of task j, ExecutionTime max ExecutionTime is the maximum execution time of any task in the collected sequence of task execution times. min This represents the minimum task execution time among the collected task execution time sequences.

[0038] For each energy consumption data point, a min-max normalization method is used to generate normalized energy consumption data, ensuring comparability between different tasks:

[0039] Among them, TotalEnergy j For the energy consumption data of task j, TotalEnergy j 'For the normalized energy consumption data of task j, TotalEnergy max TotalEnergy is the maximum value in the collected energy consumption data sequence. min It is the minimum value in the collected energy consumption data sequence.

[0040] For each current waveform sequence, the Z-score normalization method is used to form a normalized current waveform sequence, eliminating the differences in rated power between different motors.

[0041] Among them, CurrentWaveform j(t) represents the current waveform sequence of task j, CurrentWaveform j '(t) is the normalized current waveform sequence of task j, μ is the mean of the current waveform, and σ is the standard deviation of the current waveform.

[0042] Filtering and smoothing: For normalized task execution time data and normalized energy consumption data, a moving average method is used to remove occasional spikes.

[0043] For the normalized current waveform sequence, a Butterworth low-pass filter is used with a cutoff frequency set to twice the rated drive frequency to preserve the low-frequency trend of degradation correlation while suppressing high-frequency noise.

[0044] After preprocessing, the dataset corresponding to the two manipulators is obtained as required for constructing the mirror residual features:

[0045] in: TaskID is the task identifier; ArmID is the identifier for the manipulator arm. TaskType is the task type, used to label each task according to the typical actions of the semiconductor device's operating arm. Task types include gripping, moving, loading, and unloading. ExecutionTime j ', where ' is the normalized execution time of task j; Total Energy j ', represents the normalized energy consumption data for task j; , which is the normalized current waveform sequence for task j.

[0046] By preprocessing the collected multi-factor operational data, dimensional differences under different tasks and motor power conditions can be eliminated, noise interference can be reduced, and the one-to-one correspondence between the multi-factor operational data of the two manipulators at the task level can be guaranteed. This not only provides clean and comparable input data for the subsequent construction of mirror residual features, but also ensures the repeatability and reliability of the anomaly detection process.

[0047] It should be noted that each task includes normalized task execution time, normalized energy consumption data, and normalized current waveform sequence to ensure that different tasks are comparable in terms of speed, energy efficiency, and dynamics.

[0048] Step S102: Constructing the mirror residual feature vector. Based on the normalized task execution time, normalized energy consumption data, and normalized current waveform sequence, the corresponding data of the first and second operating arms are compared according to the task type and residual time window to form task duration residual features, energy consumption residual features, and current waveform residual features. The mirror residual feature vector is then constructed based on the three residual features.

[0049] In this step, the operational data of the two manipulators are first compared under the same residual time window and task type, and the results are uniformly represented as residual features. They are then combined into a mirror residual feature vector to reflect the asymmetry of the manipulators during task execution.

[0050] By constructing mirror residual features, noise interference caused by changes in the overall operating conditions of the manipulator can be effectively filtered out, highlighting the differences in execution time, energy consumption, and current waveforms of the fixed 180° opposing dual manipulators, thereby sensitively capturing signs of component degradation and providing high-quality input for subsequent prediction models.

[0051] Step S103: Prediction model construction and anomaly detection. The mirror residual feature vector is input into the prediction model. The prediction model analyzes the mirror residual feature vector of the continuous time window, generates anomaly score set, and uses a dual threshold judgment mechanism to judge the working state of the fixed 180° opposing double manipulators and obtain anomaly result data.

[0052] In this step, the prediction model uses time series features to learn and infer the mirror residual feature vector, outputs the anomaly score results corresponding to task execution time, energy consumption and current waveform, and combines a dual threshold judgment mechanism to realize the graded judgment of anomaly level.

[0053] By combining the predictive model with the dual-threshold judgment mechanism, the accuracy and robustness of anomaly detection are guaranteed, while avoiding false alarms and false negatives caused by the single threshold method. At the same time, it can classify abnormal states, providing a reliable basis for operation and maintenance prompts and diagnostic prompts.

[0054] Step S104: Operation and maintenance prompt generation. Based on the abnormal result data, an early warning prompt is output, which includes operation and maintenance prompts and abnormal diagnosis prompts.

[0055] In this step, early warning prompts are issued based on the abnormal result data from step S103, including operation and maintenance prompts and abnormal diagnosis prompts.

[0056] The maintenance prompt is as follows: When the anomaly level is Level 1, a standard warning will be issued; When the anomaly level is level 2, the message "Maintenance recommended" will be displayed. When the anomaly level is three, the message "Check immediately" will be displayed.

[0057] The abnormal diagnosis suggests: Obtain the task duration anomaly score (Score_T), energy consumption anomaly score (Score_E), and current waveform anomaly score (Score_I). When the abnormal task duration score (Score_T) is greater than or equal to the preset abnormal task duration threshold, the message "Check transmission resistance, operating arm jamming" will be displayed. When the energy consumption anomaly score (Score_E) is greater than or equal to the preset energy consumption anomaly threshold, the message "Check motor efficiency and bearing lubrication status" will be displayed. When the current waveform anomaly score (Score_I) is greater than or equal to the preset current waveform anomaly threshold, the message "Check motor driver, encoder, or load balance" will be displayed.

[0058] Set alert release rules: The maintenance alert is lifted when the overall anomaly score for R consecutive residual time windows is less than the preset deviation threshold. Here, R is a positive integer, and in this embodiment, R is 5.

[0059] Figure 2 This is a schematic diagram illustrating the process of constructing mirrored residual feature vectors according to an example embodiment of this application. For example... Figure 2 As shown, the mirror residual feature vector construction method provided in this embodiment includes: In this embodiment, to ensure the stability and comparability of each residual feature, a time window mechanism is introduced before constructing the residual features. Within the constructed residual time window, the task duration residual features, energy consumption residual features, and current waveform residual features are calculated.

[0060] Step S1021: Construct residual time windows. Within each residual time window, construct task duration sample sets, energy consumption sample sets, and current waveform sample sets based on task type. Then, aggregate each sample set according to the operator arm identifier to form the first operator arm sample set and the second operator arm sample set. Finally, start residual feature analysis when the residual analysis conditions are met.

[0061] In this step, continuous time windows are used for data aggregation. The residual time window is a division of the continuous operation process of the operating arm into fixed time lengths:

[0062] in, Let be the start time of the k-th residual time window. Rolling updates are performed with a step size W (e.g., W=10min).

[0063] Within each residual time window, tasks falling within that window are selected according to task type. The normalized task execution time, normalized energy consumption data, and normalized current waveform sequence of each task are then sampled as task duration samples, energy consumption samples, and current waveform samples, respectively, forming task duration sample sets, energy consumption sample sets, and current waveform sample sets. Each normalized current waveform sequence corresponds to a task instance and participates in the set statistics as an independent sample unit.

[0064] Subsequently, based on the operator arm identification, the task duration sample set, energy consumption sample set, and current waveform sample set are aggregated to form the first operator arm sample set and the second operator arm sample set.

[0065] Finally, the corresponding data of the first manipulator sample set and the second manipulator sample set are compared to obtain the task duration residual characteristics, energy consumption residual characteristics, and current waveform residual characteristics.

[0066] To ensure the statistical stability of the residual characteristics, this embodiment sets residual analysis conditions. Residual characteristic analysis is initiated only when these conditions are met. Specifically: When the task duration samples, energy consumption samples, and current waveform samples in the first operator arm sample set are all greater than or equal to the preset minimum sample threshold (n) min Furthermore, the task duration samples, energy consumption samples, and current waveform samples in the second manipulator sample set are all greater than or equal to the preset minimum sample threshold (n). min When the residual time window is reached, the residual characteristics of task duration, energy consumption, and current waveform are calculated. Otherwise, the residual time window is invalid and will not be included in the predictive analysis.

[0067] In this embodiment, the preset minimum sample threshold n min The value is 5-10 samples.

[0068] This embodiment maintains the temporal consistency and low latency of online anomaly detection by using a fixed time window and minimum sample constraints, while ensuring the statistical reliability of residual features and avoiding occasional false alarms caused by capacity fluctuations or task shortages.

[0069] It should be noted that, based on the structural characteristics of the fixed 180° opposing dual operating arms, this embodiment samples the normalized task execution time, normalized energy consumption data and normalized current waveform sequence under the same residual time window and task type to form corresponding task time samples, energy consumption samples and current waveform samples, which is the mirror operation. The task duration samples, energy consumption samples, and current waveform samples are collected according to the operator arm identifier to form the first operator arm sample set and the second operator arm sample set, and then compared. The difference between them is the mirror residual.

[0070] Step S1022: In the task duration sample set, calculate the median of the task duration samples of the first operating arm and the second operating arm and the absolute deviation of the median of the task duration samples, thereby constructing the task duration residual feature, and forming the standardized task duration residual through standardization processing.

[0071] In this step, within each residual time window (Win k Within each task type, based on the task duration sample sets in the first and second operator arm sample sets, the median and absolute deviation of the task duration samples for both operator arms are calculated, and task duration residual features are constructed accordingly. Specifically: All tasks satisfying the task type τ (TaskType=τ) and falling within the residual time window (Win) k The task duration sample set is {ExcutionTime} j |TaskType=τ,j∈Win k},but: The median sample duration of the first manipulator's task time was: The absolute deviation of the median of the task duration sample is ; The median sample duration of the second manipulator's task time was: The absolute deviation of the median of the task duration sample is .

[0072] Where τ is the task type, and T is the set of task duration samples belonging to task type τ.

[0073] It should be noted that the median is the middle value after the sample is sorted, and the median absolute deviation (MAD) is the median of the absolute deviation relative to the median. It is a robust statistical method in the prior art.

[0074] Then, the task duration residual features are calculated based on the median of the task duration samples:

[0075] in: τ represents the task type; , where is the median of the task duration of the first manipulator within the k-th residual time window; , which is the median of the task duration of the second manipulator within the k-th residual time window.

[0076] The task duration residual features are Z-score standardized to form standardized task duration residuals:

[0077] in: For tasks of type τ and belonging to the residual time window Win k Task duration residual characteristics within; , is the absolute deviation of the sample median of the task duration of the first manipulator; , which is the absolute deviation of the sample median of the task duration of the second manipulator.

[0078] It should be noted that using the median of the task duration sample and the absolute deviation of the median of the task duration sample aims to reduce the impact of occasional anomalies on the statistical sample, making the task duration residual features more stably reflect the systematic speed differences between the two manipulators in the same type of task. Through Z-score standardization, the task duration residual features are unified into a dimensionless space, which facilitates their combination with the standardized energy consumption residual and the standardized current waveform residual to form a mirror residual feature vector.

[0079] Step S1023: In the energy consumption sample set, calculate the median of the energy consumption samples of the first operating arm and the second operating arm and the absolute deviation of the median of the energy consumption samples, thereby constructing the energy consumption residual feature, and forming the standardized energy consumption residual through standardization processing.

[0080] In this step, within each residual time window (Win k Based on the energy consumption sample sets in the first and second operator arm sample sets, according to the task type, the median energy consumption sample and the absolute deviation of the median energy consumption sample are calculated for both operator arms, and the energy consumption residual characteristics are constructed accordingly. Specifically: All tasks satisfying the condition of task type τ (TaskType=τ) and belonging to the residual time window (Win) k The energy consumption sample set within {TotalEnergy} is... k |TaskType=τ,j∈Win k},but: The median energy consumption of the first manipulator was: The absolute deviation of the median of its energy consumption sample is ; The median energy consumption of the second manipulator was: The absolute deviation of the median of its energy consumption sample is .

[0081] Where τ is the task type, and T is the set of energy consumption samples belonging to task type τ.

[0082] It should be noted that the median is the middle value after the sample is sorted, and the median absolute deviation (MAD) is the median of the absolute deviation relative to the median. It is a robust statistical method in the prior art.

[0083] Then, the energy consumption residual characteristics are calculated based on the median of the energy consumption samples:

[0084] in: τ represents the task type; , where is the median energy consumption of the first operating arm within the k-th residual time window; , which is the median energy consumption of the second operating arm within the k-th residual time window.

[0085] The energy consumption residual is Z-score standardized to form a standardized energy consumption residual:

[0086] in: For task type τ, and belonging to the residual time window Win k Energy consumption residual characteristics within; The absolute deviation of the median energy consumption sample of the first manipulator; The absolute deviation of the sample median energy consumption of the second manipulator.

[0087] It should be noted that the use of the median energy consumption sample and the absolute deviation of the median energy consumption sample aims to reduce the impact of occasional abnormal energy consumption data (such as current fluctuations or sampling spikes in a single task) on the statistics, so that the energy consumption residual features can more stably reflect the energy consumption differences between the two operating arms under the same task. Through Z-score standardization, the energy consumption residual features are unified into a dimensionless space, which facilitates the formation of a mirror residual feature vector together with the standardized task duration residual and the standardized current waveform residual.

[0088] Step S1024: In the current waveform sample set, calculate the median and absolute deviation of the current waveform sample median of the first operating arm and the second operating arm, thereby constructing the current waveform residual features, and forming standardized current waveform residual features through standardization processing.

[0089] In this step, within each residual time window (Win k Based on the current waveform samples and according to the task type, the median and absolute deviation of the median of the current waveform samples for both operating arms are calculated, and the residual characteristics of the current waveform are constructed accordingly. Specifically: The set of all current waveform samples that satisfy the task type τ (TaskType=τ) and belong to the residual time window (Wink) is: ,but: The median of the current waveform samples of the first manipulator is The absolute bias of the median of its current waveform samples is ; The median of the current waveform samples of the second manipulator is The absolute deviation of the median of its current waveform samples is .

[0090] Where τ is the task type, and T is the set of current waveform samples belonging to task type τ.

[0091] It should be noted that the median is the middle value after the sample is sorted, and the median absolute deviation (MAD) is the median of the absolute deviation relative to the median. It is a robust statistical method in the prior art.

[0092] Then, the residual characteristics of the current waveform are calculated based on the median of the current waveform samples:

[0093] in: τ represents the task type; , is the median of the first operating arm current waveform sample within the k-th residual time window; , which is the median of the second operating arm current waveform samples within the k-th residual time window.

[0094] The residual characteristics of the current waveform are Z-score normalized to form a normalized current waveform residual:

[0095] in: For task type τ, and belonging to the residual time window Win k Current waveform residual characteristics within; , which is the absolute deviation of the median of the current waveform samples of the first operating arm; , which is the absolute deviation of the median of the current waveform samples of the second operating arm.

[0096] In this embodiment, the standardized current waveform residual characteristics are used to comprehensively characterize the differences between the two manipulators in terms of speed, energy efficiency, and dynamic behavior.

[0097] It should be noted that using the median and absolute deviation of the current waveform samples aims to reduce the impact of outliers caused by occasional interference (such as transient electrical noise and sampling jitter) in a single current waveform, so that the current waveform residual features can more robustly reflect the dynamic differences between the two manipulators under similar tasks. Z-score standardization unifies the current waveform residual features into a dimensionless space, facilitating their combination with the standardized task duration residual and standardized energy consumption residual to form a mirror residual feature vector.

[0098] Step S1025: Under the same residual time window and task type, the standardized task duration residual, standardized energy consumption residual, and standardized current waveform residual are jointly constructed into a mirror residual feature vector.

[0099] In this step, based on the standardized task duration residual, standardized energy consumption residual, and standardized current waveform residual, a mirror residual feature vector is constructed according to the task type within each residual time window:

[0100] in: WinID is the identifier for the residual time window; TaskType is the task type; ZT represents the residual time of standardized task execution. ZE represents the standardized energy consumption residual. ZI represents the standardized current waveform residual.

[0101] Within the k-th residual time window, when the task type is τ, its mirror residual feature vector is specifically represented as follows:

[0102] It should be noted that by jointly constructing a feature vector from the three types of standardized residuals under a unified time window and task type, the differences between the two manipulators in three dimensions—speed (task duration), energy efficiency (energy consumption), and dynamic behavior (current waveform)—can be reflected simultaneously, thereby enhancing the comprehensiveness and stability of degradation judgment. This mirrored residual feature vector serves as the input to the prediction model, providing multi-dimensional support for the subsequent dual-threshold judgment mechanism.

[0103] Figure 3 This is a schematic diagram illustrating the process of predictive model construction and anomaly detection according to an example embodiment of this application. For example... Figure 3 As shown, the prediction model construction and anomaly detection method provided in this embodiment includes: Step S1031: Construct a prediction model based on the time series characteristics of the mirror residual feature vector.

[0104] In this embodiment, the prediction model is a deep neural network model, and lightweight optimization is performed using parameter pruning and 8-bit integer quantization. Specifically, the prediction model employs a Gated Recurrent Unit (GRU) network, a simplified structure of recurrent neural networks. This model features a small number of parameters and high computational efficiency, making it suitable for processing sequences of mirror residual feature vectors over continuous time windows, thereby extracting the evolution of residuals over time.

[0105] The prediction model can be implemented based on existing open-source frameworks (such as TensorFlow Lite, PyTorch Mobile, etc.). This solution does not involve the model training process, but only performs lightweight optimizations during implementation, including: Parameter pruning: By analyzing the sensitivity of each weight in the prediction model to the output, the coefficients of neurons with low contribution or convolution kernels are set to zero, thereby reducing computational and storage requirements. In implementation, a sparse weight matrix can be automatically generated using model compression tools (such as the TensorFlow Model Optimization Toolkit), and then a sparse computation library can be deployed on edge devices for efficient inference. In this embodiment, parameter pruning is used to reduce redundant computations of the prediction model when running on edge servers or local gateways, ensuring real-time processing of mirrored residual feature vectors at the window-level temporal granularity.

[0106] The 8-bit integer quantization method: While maintaining the original prediction model structure, this method compresses floating-point weights and activation values ​​into 8-bit integer representations, significantly reducing the memory footprint and computational overhead of the prediction model. In implementation, the quantization conversion interface provided by PyTorch Mobile or TensorFlow Lite can be used to complete weight quantization offline, and the quantization inference kernel can be invoked at the edge for execution. In this embodiment, the 8-bit integer quantization method is used to further compress model weights and computational precision, reducing storage and computational requirements, enabling the prediction model to run stably under the limited computing power of edge servers or local gateways, meeting the real-time requirements of online anomaly detection.

[0107] Step S1032: The mirror residual feature vector, composed of the standardized task execution time residual, the standardized energy consumption residual, and the standardized current waveform residual, is used as the input of the prediction model to predict and analyze the working state of the fixed 180° opposing dual operating arms, and outputs a multi-dimensional anomaly score set.

[0108] In this step, the mirror residual feature vector output in step S1025 is used as the input to the prediction model to obtain an anomaly score set. The anomaly score set includes component-level anomaly scores and a comprehensive anomaly score. Wherein: Component-level anomaly scores: These are the task duration anomaly scores (Score_T), energy consumption anomaly scores (Score_E), and current waveform anomaly scores (Score_I) output by the prediction model based on the standardized task duration residual (ZT), standardized energy consumption residual (ZE), and standardized current waveform residual (ZI) in the mirror residual feature vector, respectively. The component-level anomaly scores are used to characterize the specific deviations in each dimension, providing a detailed basis for subsequent operation and maintenance prompts.

[0109] Comprehensive Anomaly Score: The comprehensive anomaly score (Score) is obtained by weighting the task duration anomaly score (Score_T), energy consumption anomaly score (Score_E), and current waveform anomaly score (Score_I).

[0110] Wherein, β1, β2 and β3 are the components of each weight, and must satisfy β1+β2+β3=1.

[0111] In this embodiment, the comprehensive anomaly score will provide the data foundation for the dual threshold mechanism.

[0112] Step S1033: Based on the abnormal score set, an abnormal judgment is made on the working state of the fixed 180° opposing dual operating arms by adopting a dual threshold judgment mechanism that combines the deviation threshold and the confidence threshold.

[0113] In this step, a dual-threshold determination is performed on the abnormal score set to reduce false alarms. The dual-threshold determination mechanism includes: When the comprehensive anomaly score is greater than or equal to the preset emergency threshold, the operating arm is determined to be abnormal and the anomaly level is marked as level three. When the overall anomaly score is less than the preset emergency threshold and greater than or equal to the preset deviation threshold, it is determined that there is an abnormal deviation in the current residual time window, and the continuous counter is incremented by 1. When the comprehensive anomaly score shows abnormal deviations in N consecutive residual time windows, the operating arm is determined to be abnormal and the anomaly level is marked as Level 1 anomaly. When the comprehensive anomaly score shows abnormal deviations in all M consecutive residual time windows, the operating arm is determined to be abnormal and the anomaly level is identified as Level 2 anomaly. When the overall anomaly score is less than the preset deviation threshold, the continuous counter is reset to zero.

[0114] Where N and M are both positive integers, and M is greater than N, and the emergency threshold is greater than the deviation threshold. In this embodiment, N is 3 and M is 10.

[0115] The above-mentioned dual-threshold judgment mechanism can ensure that the anomaly detection results take into account not only the magnitude of the difference, but also the stability of the evidence, thereby effectively reducing the false alarm rate.

[0116] When an anomaly is detected, anomaly result data is constructed:

[0117] in: TriggerWinID represents the residual time window in which the abnormal deviation occurred; ContextWinIDSet represents the consecutive residual time windows where the anomaly occurred; TaskType is the task type; AnonalyLevel is the anomaly level, including Level 1 anomaly, Level 2 anomaly, and Level 3 anomaly. ScoreSet is a set of anomaly scores, including task duration anomaly score (Score_T), energy consumption anomaly score (Score_E), current waveform anomaly score (Score_I), and comprehensive anomaly score (Score).

[0118] Record abnormal results data to the log and generate log records.

[0119] It should be noted that the TriggerWinID, ContextWinIDSet, and TaskType fields are used for tracing and logging in the project implementation to facilitate subsequent analysis of the context window range in which the anomaly occurred, and are not used as the basis for subsequent warning prompts.

[0120] Figure 4 This is a schematic diagram illustrating the structure of a component maintenance prediction system for a semiconductor device according to an example embodiment of this application. Figure 4 As shown, the semiconductor equipment component maintenance prediction system 400 provided in this embodiment includes: a multi-factor operation data acquisition and preprocessing module 410, a mirror residual feature vector construction module 420, a prediction model construction and anomaly detection module 430, and an operation and maintenance prompt generation module 440.

[0121] The multi-factor operation data acquisition and preprocessing module 410 acquires multi-factor operation data of the fixed 180° opposing double operating arms during equipment operation, including task execution time, energy consumption data and current waveform sequence, and performs normalization processing on the multi-factor operation data to form normalized task execution time, normalized energy consumption data and normalized current waveform sequence. The mirror residual feature vector construction module 420 compares the corresponding data of the first operating arm and the second operating arm according to the normalized task execution time, normalized energy consumption data and normalized current waveform sequence, according to the task type and residual time window, to form task duration residual features, energy consumption residual features and current waveform residual features, and jointly constructs a mirror residual feature vector based on the three residual features. The prediction model construction and anomaly detection module 430 inputs the mirror residual feature vector into the prediction model. The prediction model analyzes the mirror residual feature vector of the continuous time window, generates anomaly score set, and uses a dual threshold judgment mechanism to judge the working state of the fixed 180° opposing double operating arms to obtain anomaly result data. The operation and maintenance prompt generation module 440 outputs early warning prompts based on the abnormal result data. The early warning prompts include operation and maintenance prompts and abnormal diagnosis prompts.

[0122] Figure 5 This is a schematic diagram of the structure of an electronic device according to an example embodiment of this application. For example... Figure 5 As shown, the electronic device 500 provided in this embodiment includes: a processor 501 and a memory 502; wherein: Memory 502 is used to store computer programs, and the memory may also be flash memory.

[0123] Processor 501 is used to execute the execution instructions stored in the memory to implement the various steps in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0124] Alternatively, the memory 502 can be either standalone or integrated with the processor 501.

[0125] When the memory 502 is a device independent of the processor 501, the electronic device 500 may further include: Bus 503 is used to connect the memory 502 and the processor 501.

[0126] This embodiment also provides a readable storage medium storing a computer program, which, when executed by at least one processor of an electronic device, enables the electronic device to perform the methods provided in the various embodiments described above.

[0127] This embodiment also provides a program product including a computer program stored in a readable storage medium. At least one processor of an electronic device can read the computer program from the readable storage medium, and the at least one processor executes the computer program to cause the electronic device to perform the methods provided in the various embodiments described above.

[0128] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the foregoing claims.

[0129] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for predicting the maintenance of components in semiconductor equipment, characterized in that, include: Multi-factor operation data acquisition and preprocessing: During equipment operation, multi-factor operation data of fixed 180° opposing double operating arms are collected, including task execution time, energy consumption data and current waveform sequence. The multi-factor operation data is then normalized to form normalized task execution time, normalized energy consumption data and normalized current waveform sequence. The mirror residual feature vector is constructed by comparing the corresponding data of the first and second operating arms according to the normalized task execution time, normalized energy consumption data and normalized current waveform sequence, according to the task type and residual time window, to form task duration residual features, energy consumption residual features and current waveform residual features, and then constructing the mirror residual feature vector based on the three residual features. The prediction model is constructed and anomaly detection is performed by inputting the mirror residual feature vector into the prediction model. The prediction model analyzes the mirror residual feature vector of the continuous time window, generates anomaly score set, and uses a dual threshold judgment mechanism to judge the working state of the fixed 180° opposing double manipulators and obtain anomaly result data. Operation and maintenance prompts are generated, and early warning prompts are output based on the abnormal result data. The early warning prompts include operation and maintenance prompts and abnormal diagnosis prompts.

2. The component maintenance prediction method for semiconductor equipment according to claim 1, characterized in that, The construction of the mirror residual feature vector includes: Construct a residual time window, and within each residual time window, construct a task duration sample set, an energy consumption sample set, and a current waveform sample set based on the task type. Then, aggregate each sample set according to the operator arm identifier to form the first operator arm sample set and the second operator arm sample set. Finally, start residual feature analysis when the residual analysis conditions are met. In the task duration sample set, the median of the task duration samples of the first operating arm and the second operating arm and the absolute deviation of the median of the task duration samples are calculated to construct the task duration residual features, and standardized task duration residuals are formed through standardization processing. In the energy consumption sample set, the median of the energy consumption samples of the first operating arm and the second operating arm and the absolute deviation of the median of the energy consumption samples are calculated to construct the energy consumption residual characteristics, and standardized energy consumption residuals are formed through standardization processing. In the current waveform sample set, the median and absolute deviation of the current waveform sample of the first operating arm and the second operating arm are calculated to construct the current waveform residual features, and standardized current waveform residual features are formed through standardization processing. Under the same residual time window and task type, the standardized task duration residual, standardized energy consumption residual, and standardized current waveform residual are jointly constructed into a mirror residual feature vector.

3. The component maintenance prediction method for semiconductor equipment according to claim 1, characterized in that, The prediction model construction and anomaly detection include: A prediction model is constructed based on the time series characteristics of mirror residual feature vectors; The mirror residual feature vector, composed of standardized task execution time residual, standardized energy consumption residual, and standardized current waveform residual, is used as the input of the prediction model to predict and analyze the working state of the fixed 180° opposing dual operating arms, and outputs a multi-dimensional set of anomaly scores. Based on the aforementioned abnormal score set, a dual-threshold judgment mechanism combining deviation threshold and confidence threshold is adopted to determine the abnormal working state of the fixed 180° opposing dual manipulators.

4. The component maintenance prediction method for semiconductor equipment according to claim 1, characterized in that, The residual time window is a division of the continuous operation process of the operating arm into fixed time lengths; Within each residual time window, tasks falling into the residual time window are selected according to task type. The normalized task execution time, normalized energy consumption data and normalized current waveform sequence of each task are sampled as task duration sample, energy consumption sample and current waveform sample, respectively, to form a task duration sample set, energy consumption sample set and current waveform sample set. Based on the operator arm identification, the task duration sample set, energy consumption sample set, and current waveform sample set are collected to form the first operator arm sample set and the second operator arm sample set. By comparing the corresponding data of the first manipulator sample set with the second manipulator sample set, the residual characteristics of task duration, energy consumption, and current waveform are obtained.

5. The component maintenance prediction method for semiconductor equipment according to claim 1, characterized in that, The component maintenance prediction method is based on the structural characteristics of a fixed 180° opposing double operating arm. Under the same residual time window and task type, the normalized task execution time, normalized energy consumption data and normalized current waveform sequence are sampled to form corresponding task time samples, energy consumption samples and current waveform samples, which is the mirror operation. The task duration samples, energy consumption samples, and current waveform samples are collected according to the operator arm identifier to form the first operator arm sample set and the second operator arm sample set, and then compared. The difference between them is the mirror residual.

6. The component maintenance prediction method for semiconductor equipment according to claim 1, characterized in that, The dual threshold determination mechanism includes: When the comprehensive anomaly score is greater than or equal to the preset emergency threshold, the operating arm is determined to be abnormal and the anomaly level is marked as level three. When the overall anomaly score is less than the preset emergency threshold and greater than or equal to the preset deviation threshold, it is determined that there is an abnormal deviation in the current residual time window, and the continuous counter is incremented by 1. When the comprehensive anomaly score shows abnormal deviations in N consecutive residual time windows, the operating arm is determined to be abnormal and the anomaly level is marked as Level 1 anomaly. When the comprehensive anomaly score shows abnormal deviations in all M consecutive residual time windows, the operating arm is determined to be abnormal and the anomaly level is identified as Level 2 anomaly. When the overall anomaly score is less than the preset deviation threshold, the continuous counter is reset to zero. Where N and M are both positive integers, and M is greater than N, and the emergency threshold is greater than the deviation threshold.

7. The component maintenance prediction method for semiconductor equipment according to claim 1, characterized in that, The prediction model is a deep neural network model, and it is optimized using parameter pruning and 8-bit integer quantization.

8. A component maintenance prediction system for semiconductor equipment, characterized in that, include: The multi-factor operation data acquisition and preprocessing module acquires multi-factor operation data of the fixed 180° opposing double operating arms during equipment operation, including task execution time, energy consumption data and current waveform sequence, and performs normalization processing on the multi-factor operation data to form normalized task execution time, normalized energy consumption data and normalized current waveform sequence. The mirror residual feature vector construction module compares the corresponding data of the first and second operating arms according to the normalized task execution time, normalized energy consumption data and normalized current waveform sequence, according to the task type and residual time window, to form task duration residual features, energy consumption residual features and current waveform residual features, and jointly constructs a mirror residual feature vector based on the three residual features. The prediction model construction and anomaly detection module inputs the mirror residual feature vector into the prediction model. The prediction model analyzes the mirror residual feature vector of the continuous time window, generates anomaly score set, and uses a dual threshold judgment mechanism to judge the working state of the fixed 180° opposing double manipulators and obtain anomaly result data. The operation and maintenance prompt generation module outputs early warning prompts based on the abnormal result data. The early warning prompts include operation and maintenance prompts and abnormal diagnosis prompts.

9. An electronic device, characterized in that, include: processor; as well as, Memory for storing the executable instructions of the processor; The processor is configured to execute the method of any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 7.

Images