A method for semiconductor equipment chamber maintenance group decision making

Abstract

The application discloses a kind of semiconductor equipment chamber maintenance group decision-making methods, it is related to semiconductor equipment maintenance technical field;The method collects process parameter multidimensional time series data and other multi-source data, constructs dynamic heterogeneous correlation diagram with process chamber node as core, and based on component sharing association automatically deduces the failure cascade association between chambers and its propagation intensity coefficient, the unified characterization of the physical coupling relationship of multiple chambers is carried out;Adopt double-channel graph time sequence coding and cumulative risk function estimation joint prediction framework, output each chamber's survival probability curve and export remaining reliable operation time length estimate value and failure risk score;Construct hierarchical multi-agent collaborative decision-making framework, output the scheduling decision used to arrange equipment chamber maintenance job;The application can accurately characterize the evolution law of equipment health status under the complex coupling relationship of multiple chambers, realize the global optimization matching of maintenance scheduling and production task under the premise of guaranteeing the physical feasibility of maintenance scheme, reduce unplanned downtime loss.

Description

Technical Field

[0001] This invention relates to the field of semiconductor equipment maintenance technology, and more specifically to a semiconductor equipment chamber maintenance group decision-making method. Background Technology

[0002] The stable operation of semiconductor manufacturing equipment is a key factor in ensuring product yield and production line efficiency. As process technologies continue to evolve towards smaller nodes, the complexity of core process equipment such as etching and thin film deposition continues to increase. A single piece of equipment typically contains multiple process chambers, which are interconnected through shared gas pipelines, RF power supplies, vacuum systems, and other components, forming complex physical coupling relationships. An abnormal state in any chamber can propagate to adjacent chambers through shared components or process contamination, triggering cascading failures and causing unplanned downtime of the equipment cluster, resulting in significant economic losses.

[0003] Currently, the industry's prevailing equipment maintenance strategies still primarily rely on scheduled maintenance and reactive repair. Scheduled maintenance is performed based on fixed time intervals or the number of pieces processed, without considering the differences in the actual operating conditions and health status of each chamber, easily leading to both under-maintenance and over-maintenance. Reactive repair, on the other hand, intervenes only after a failure has occurred, by which time production interruptions and potential product quality impacts have already taken place. Neither model can achieve a precise match between maintenance timing and production tasks.

[0004] In recent years, data-driven predictive maintenance methods have gradually attracted attention. Some studies utilize time-series data from equipment sensors, combined with deep learning models, to predict the remaining service life or failure probability of a single chamber, thus improving the rationality of maintenance timing to some extent. However, such methods have the following shortcomings: First, most methods model each chamber as an independent entity, failing to fully consider the complex coupling relationships between chambers due to shared components, process correlations, and fault propagation, leading to deviations between the prediction results and the actual operating status of the equipment cluster. Second, there is a lack of an effective global coordination mechanism between the prediction results and maintenance decisions. When multiple chambers generate maintenance needs simultaneously, it is difficult to balance capacity and resolve conflicts at the equipment cluster level. Third, maintenance decisions are not coupled with the physical constraints unique to semiconductor equipment, such as vacuum recovery, clean recovery, wafer capacity, and continuous processing batches. The generated maintenance timings are often physically unexecutable, making it difficult to guarantee the global optimality and feasibility of the maintenance schedule.

[0005] Therefore, there is an urgent need for an intelligent maintenance method for semiconductor equipment chambers that can fully characterize the complex coupling relationships of equipment clusters and, based on this, combine the physical constraints of semiconductor equipment to achieve global collaborative optimization decisions. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, this invention aims to provide a semiconductor equipment chamber maintenance group decision-making method. This method solves the technical problems that existing predictive maintenance methods cannot accurately characterize the complex coupling relationships between multiple chambers within a semiconductor equipment cluster due to component sharing and fault cascading, and lack a global maintenance scheduling optimization decision-making mechanism that combines component collaboration, production priorities, and semiconductor physical constraints at the equipment cluster level.

[0007] To solve the above problems, the present invention adopts the following technical solution: A semiconductor device chamber maintenance group decision-making method includes: By collecting multi-dimensional time-series data of process parameters through sensors deployed in the semiconductor equipment chamber, and collecting equipment maintenance history data, process specification definition data, semiconductor processing scheduling data and equipment static attribute data, a standardized multi-source dataset is obtained through standardization. Based on the standardized multi-source dataset, six types of entity nodes are defined: process chamber nodes, component nodes, fault event nodes, maintenance operation nodes, process recipe nodes, and production work order nodes. According to six types of association relationships between entity nodes—composition association, fault attribution association, maintenance response association, process adaptation association, component sharing association, and fault cascading association—a dynamic heterogeneous association graph is constructed by combining the timestamp attribute, effective time interval attribute, and data sampling granularity attribute of each association relationship. Specifically, fault cascading associations are automatically derived based on the component sharing association between two process chamber nodes, and the fault propagation risk is characterized by a fault propagation intensity coefficient. Based on the dynamic heterogeneous correlation graph, a joint prediction framework of dual-channel graph temporal coding and cumulative risk function estimation is adopted. After fusing dynamic runtime sequence features and topological neighborhood structure features, the survival probability curve of each process chamber node is output, and the estimated value of the remaining reliable runtime is derived from the survival probability curve. At the same time, the type of failure event with the highest probability of occurrence and its risk score are output. A hierarchical multi-agent collaborative decision-making framework is constructed, mapping each process chamber to a chamber maintenance agent and setting a global maintenance arbitration agent. Before generating a maintenance request, the chamber maintenance agent performs physical constraint verification and maintenance window feasibility prediction based on vacuum recovery, cleanroom recovery, wafer capacity and continuous processing batch constraints. The global maintenance arbitration agent resolves conflicts based on all maintenance requests and the global production status, and outputs a maintenance scheduling decision for arranging equipment chamber maintenance operations.

[0008] The beneficial effects of this invention are as follows: First, by constructing a dynamic heterogeneous association graph, a unified explicit model is used to model various coupling relationships among process chamber nodes within the equipment cluster, such as compositional associations, component sharing associations, fault cascading associations, and process adaptation associations. In particular, fault cascading associations are automatically derived based on component sharing associations, and the risk of fault propagation is quantified by the propagation intensity coefficient. Compared with the modeling method in the prior art that treats each chamber as an independent individual, this method can more accurately characterize the real topology and failure propagation path of the semiconductor equipment cluster, providing richer structural prior information for subsequent health status prediction.

[0009] Second, a joint prediction framework combining dual-channel graph temporal coding and cumulative risk function estimation is adopted. The temporal self-attention coding channel extracts the runtime sequence dependency features of each chamber, while the graph attention coding channel extracts the topological neighborhood structure features and distinguishes the influence intensity of different association types. The cross-attention mechanism is used to achieve deep fusion of the two types of features. This framework can simultaneously capture the temporal evolution of the chamber's own operating state and the coupling influence exerted by adjacent chambers through shared components and fault propagation paths, thereby improving the accuracy and robustness of remaining reliable runtime estimation and fault event risk ranking.

[0010] Third, by constructing a hierarchical multi-agent collaborative decision-making framework and embedding feasibility verification based on physical constraints of semiconductor equipment such as vacuum recovery, clean recovery, wafer capacity, and continuous processing batches into maintenance decisions, as well as maintenance window feasibility prediction and component collaborative maintenance optimization, it is possible to effectively resolve time window conflicts and resource competition among concurrent maintenance requests in multiple chambers, reduce unplanned downtime losses and component replacement costs, and achieve global optimization matching between maintenance scheduling and production tasks, while ensuring the physical feasibility of the maintenance plan. Attached Figure Description

[0011] Figure 1 This is a flowchart of the method of the present invention.

[0012] Figure 2 This is a schematic diagram of a dynamic heterogeneous correlation graph.

[0013] Figure 3 This is a structural diagram of the joint prediction framework.

[0014] Figure 4 This is a diagram of a hierarchical multi-agent collaborative decision-making framework.

[0015] Figure 5 To output a maintenance scheduling decision flowchart. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to specific embodiments. It should be noted that these embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Simple improvements to the method under the premise of the present invention are all within the scope of protection claimed by the present invention. For ease of description, "maintenance" and "maintenance operation" in this document refer to health restoration actions performed on process chambers or components, including preventive maintenance and emergency maintenance.

[0017] See Figure 1 This is a semiconductor device chamber maintenance group decision-making method, including steps S100 to S400.

[0018] S100. Collect multi-dimensional time-series data of process parameters by sensors deployed in the semiconductor equipment chamber, and collect equipment maintenance history data, process specification definition data, semiconductor processing scheduling data and equipment static attribute data, and standardize them to obtain a standardized multi-source dataset.

[0019] The multi-dimensional time-series data of process parameters are collected in real time by sensors deployed on the process chamber and its components. Standardization includes: performing timestamp alignment, missing value imputation and Z-score standardization on the multi-dimensional time-series data of process parameters; performing one-hot encoding conversion on the classification field; and uniformly converting the time field to Unix timestamps and normalizing it based on hours to form a standardized multi-source dataset.

[0020] Equipment maintenance history data includes: maintenance type, maintenance start and end time, replacement parts list, anomaly cause analysis, maintenance personnel information, maintenance work order status, pure maintenance operation time, maintenance cost details, anomaly level, anomaly impact range, post-maintenance equipment operating parameters, and maintenance acceptance results.

[0021] The process specification definition data includes: process specification number, process type, core process parameter set, compatible substrate type, production batch number, process specification version number, process specification effective time, process specification expiration time, compatible chamber list, process yield index, and process stability parameters; the core process parameter set includes power, temperature, gas flow rate, and etching time.

[0022] Semiconductor processing scheduling data includes: work order number, planned number of wafers to be processed, delivery deadline, process path planning, task priority, actual number of wafers processed, task start time, task completion time, task status, associated batch number, compatible equipment list, quality inspection standards, and delay penalty rules.

[0023] The static attribute data of the equipment includes: chamber number, equipment model, list of shared components of the chamber, rated operating parameters of the chamber, chamber commissioning time, most recent calibration time of the chamber, maintenance history of the chamber, chamber component ledger, production line to which the chamber belongs, responsible personnel of the chamber, chamber location number, cumulative operating years of the chamber, chamber depreciation rate, vacuum system constraint parameters, cleanliness constraint parameters, and transmission reference parameters. Among them, the rated operating parameters of the chamber include the rated power of the chamber, the rated pressure of the chamber, and the rated temperature of the chamber; the vacuum system constraint parameters include the standard vacuuming time after vacuum breaking of each chamber, the vacuum leak detection verification time, and the maximum number of chambers that can be vacuumed in parallel in the shared vacuum pipeline; the cleanliness constraint parameters include the nitrogen purging time after opening the cover for maintenance, the particle concentration compliance verification time, and the maximum chamber quota for simultaneous opening operations in the clean area; the transmission reference parameters include the rated number of wafers per chamber, the total capacity of the matching Loadlock buffer chamber, and the minimum continuous processing batch requirement.

[0024] The multi-dimensional time-series data of process parameters include: RF reflected power, chamber pressure, chamber temperature, fluorine flow rate, oxygen flow rate, argon flow rate, RF power output, plasma density, bias voltage, helium leakage rate, vacuum pump speed, vacuum level, electrostatic chuck voltage, coolant flow rate, coolant temperature, process time deviation, ion energy, fluorine-oxygen gas pressure ratio, RF matching impedance, wafer transfer speed, chamber wall temperature, fluorine purity, oxygen purity, bias current, plasma emission spectrum intensity, vacuum pump exhaust pressure, electrostatic chuck adsorption force, total process gas flow rate, RF power reflection coefficient, chamber cleanliness index, coolant pressure, and wafer positioning accuracy, totaling 32-dimensional time-series data.

[0025] S200. Constructing a Dynamic Heterogeneous Association Graph Based on Standardized Multi-Source Datasets: Based on standardized multi-source datasets, six types of entity nodes are defined: process chamber nodes, component nodes, fault event nodes, maintenance operation nodes, process recipe nodes, and production work order nodes. A dynamic heterogeneous association graph is constructed based on six types of associations between entity nodes: composition association, fault attribution association, maintenance response association, process adaptation association, component sharing association, and fault cascading association. This is combined with the timestamp attribute, effective time interval attribute, and data sampling granularity attribute of each association. Specifically, based on the component sharing association between two process chamber nodes, the fault cascading association between the two process chamber nodes is automatically derived, and the fault propagation risk is characterized by a fault propagation intensity coefficient. The fault propagation intensity coefficient increases as the time difference between the most recent maintenance operation of the two process chamber nodes increases.

[0026] The definitions and attributes of the six types of entity nodes are as follows: The process chamber node is the core object for maintaining group decision-making. Its attributes include: process chamber number, process chamber equipment model, process chamber rated operating parameters, process chamber commissioning time, process chamber equipment number, baseline vacuuming time, cleanroom recovery time, vacuum pipeline group, cleanroom zone, and associated Loadlock capacity. Example: Process chamber node-CH-A03, plasma etching equipment P-Etch-500, chamber pressure rated range 0.5-20 mTorr, August 20, 2022, Cluster-Tool-02.

[0027] Component nodes are the constituent units of a process chamber. Their attributes include: component number, component model and specifications, component function description, process chamber to which the component belongs, and common identifier for the component. Example: Component node-CP-205, high-frequency radio frequency power supply RF-800W type, plasma excitation and power output control, process chamber to which it belongs CH-A03 / CH-A05, common identifier for the component is yes.

[0028] A fault event node is an abnormal state that affects the operation of a process chamber or component. Its attributes include: fault event number, fault event type, fault event level, fault event occurrence time, and fault event impact range. Example: Fault event node-FLT-071, abnormal particulate matter contamination in the chamber, high severity level, March 15, 2024, 09:22, affecting process chambers CH-A03 / CH-A05.

[0029] Maintenance operation nodes are health recovery actions performed on process chambers or components. Their attributes include: maintenance operation order number, maintenance operation type, maintenance operation execution time, pure maintenance operation time, maintenance operation replacement parts list, whether vacuum is broken, and whether the clean environment is damaged. Example: Maintenance operation node-WO-8812, preventive maintenance, March 15, 2024, 14:00, 150 minutes, component replacement node-CP-205 high frequency radio frequency power supply.

[0030] A process recipe node is a set of process parameters for production execution. Its attributes include: process specification number, process type, core parameter set, and list of applicable process chambers. Example: Process recipe node-PR-330, polysilicon gate plasma etching process, RF power 350 watts, etching time 45 seconds, chamber pressure 5 millitor, applicable process chambers CH-A03 / CH-A05 / CH-A07.

[0031] A production work order node is a semiconductor processing task undertaken by a process chamber. Its attributes include: work order number, planned number of wafers to be processed, process path, and delivery deadline. Example: Production work order node-LOT-5621, 50 300mm substrates, PR-330 etching process to PR-412 cleaning process, March 18, 2024, 23:59.

[0032] The six types of relationships are defined as follows: Composition association represents the physical subordination relationship between process chamber nodes and component nodes, that is, a process chamber is composed of several component nodes.

[0033] Fault attribution association represents the abnormal triggering relationship between process chamber nodes or component nodes and fault event nodes, that is, a certain process chamber or component has experienced a certain fault event.

[0034] Maintenance response association represents the health recovery relationship between process chamber nodes or component nodes and maintenance operation nodes, that is, a certain maintenance operation was performed on a certain process chamber or component.

[0035] Process adaptation association represents the process execution mapping relationship between process chamber nodes and process recipe nodes, that is, a certain process chamber adapts to and executes a certain process recipe.

[0036] Component sharing association represents the shared component dependency relationship between two process chamber nodes, that is, the two process chambers share the same component.

[0037] Fault cascading association represents the abnormal propagation relationship between two process chamber nodes, that is, a fault event in one process chamber propagates to affect another process chamber through shared components or other coupling paths.

[0038] The time-series attributes include timestamp attributes, valid time interval attributes, and data sampling granularity attributes. Timestamp attributes are used to mark the occurrence time of instantaneous events, such as the fault event occurrence time 2024-03-15 09:22:00; valid time interval attributes are used to mark the start and end time range of continuous events, such as the maintenance operation execution interval [2024-03-15 14:00, 2024-03-15 16:30]; data sampling granularity attributes are used to mark the frequency of data collection or recording, such as process parameter multidimensional time-series data at the second level and maintenance history data at the hour level.

[0039] See Figure 2 This embodiment provides a feasible way to construct a dynamic heterogeneous association graph: Step S210: Extract triples from the multi-source data in the standardized multi-source dataset. The triple format is uniformly (subject entity node, association, object entity node, time-series attribute). For structured data sources, triples are directly generated using field mapping rules, mapping each row in the database table to an entity node, foreign key association fields to corresponding associations, and time fields to time-series attributes. For semi-structured data sources, a pre-trained named entity recognition and relation extraction model is used. The labeled anomaly description text training set is imported into the BERT-CRF sequence labeling model, which identifies entity references in the text and links them to predefined entity nodes. Simultaneously, associations are extracted, and time representations are extracted as time-series attributes. The extraction results are sampled and verified by operations personnel at a rate of no less than 10% per batch. For multi-dimensional time-series data sources of process parameters, an abnormal state triplet extraction method based on a sliding window is adopted: the time window size is set to 60 seconds and the sliding step size is 30 seconds. The mean value of each sensor channel data within each window is calculated. If the window mean value of any channel exceeds the preset threshold of that channel, the window is determined to be an abnormal window, and a triplet is generated with the subject being the corresponding process chamber node, the association being fault attribution association, and the object being the fault event node. The fault event number is automatically generated in the format of "FLT-Time Sequence Abnormality-Year Month Day Hour Minute Second". The fault event type is filled in as process parameter offset abnormality. The fault event level is determined according to the threshold deviation range (medium severity when the deviation does not exceed 50% of the threshold, and high severity when it exceeds 50%). The fault event occurrence time is filled in as the start timestamp of the window, and the fault event impact range is filled in as the corresponding process chamber node number.

[0040] Step S220: Create entity nodes and populate attributes for the extracted triples. All extracted entity node instances are deduplicated and aggregated. After being categorized by entity node type, nodes are created in batches using a graph database client. A globally unique node identifier is set for each entity node, with the encoding format "entity node type-entity number". If the node identifier already exists, creation is skipped, and only attribute updates are performed.

[0041] Step S230: Establish association relationships and bind temporal attributes to the extracted triples. Based on the extracted temporal triples, directed edges between entity nodes are established in batches through the relationship management interface of the graph database client. Each edge corresponds to an association relationship type and its temporal attributes are bound simultaneously: for association relationships with valid time interval attributes, their start and end timestamps are stored; for instantaneous association relationships with timestamp attributes, their occurrence time is stored; and for continuous association relationships with data sampling granularity attributes, their update frequency is stored.

[0042] Step S240: Perform fault cascading association derivation and supplementation on the extracted triples. Based on the established component-sharing associations, automatically derive and supplement fault cascading associations: When process chamber node X and process chamber node Y have a component-sharing association through component node Z, a fault cascading association is automatically established between X and Y, and a fault propagation intensity coefficient is set. The formula for calculating the fault propagation intensity coefficient is 1 / (1+e^(-α·ΔT)), where ΔT is the time difference between the most recent maintenance operations of the two process chambers, and α is the attenuation rate parameter, ranging from 0.1 to 0.5. When ΔT approaches 0, the propagation intensity coefficient approaches 0.5; when ΔT increases, the propagation intensity coefficient gradually approaches 1, indicating that the fault propagation risk increases with the increase of the maintenance time interval. The physical basis of this derivation mechanism is that when two chambers share the same component, if the time difference between their last maintenance is larger, the wear difference of the shared component between the two chambers is more significant, and the component abnormality in one chamber is more easily transmitted to the other chamber through the shared component.

[0043] Step S250: Construct a graph index. Index the retrieval fields, including process chamber number, fault event number, and the timestamp, valid time interval, and data sampling granularity attributes of the relationships, to improve the computational efficiency of subsequent graph queries and graph traversals.

[0044] S300. Based on the dynamic heterogeneous correlation graph, a joint prediction framework of dual-channel graph temporal coding and cumulative risk function estimation is used for prediction: Based on the dynamic heterogeneous correlation graph, a joint prediction framework of dual-channel graph temporal coding and cumulative risk function estimation is used to extract dynamic runtime sequence features and topological neighborhood structure features for each process chamber node. After fusion through a cross-attention mechanism, the survival probability curve of each process chamber node in the future preset time period is output. The time point corresponding to the first drop in survival probability to the preset probability threshold is found from the survival probability curve. The duration corresponding to the time point is used as the estimated value of the remaining reliable runtime of the process chamber node, and the fault event type with the highest probability of occurrence and its risk score for each process chamber node are output.

[0045] See Figure 3 The dual-channel graph temporal coding and cumulative risk function estimation joint prediction framework includes a multi-dimensional feature construction layer, a temporal self-attention coding channel, a graph attention coding channel, a cross-attention fusion layer, and a dual-task prediction output layer.

[0046] (1) Multidimensional feature construction layer, used to construct static attribute feature vectors, extend dynamic running feature matrix and topological structure feature vectors for each entity node based on dynamic heterogeneous association graph.

[0047] Static attribute feature vectors are extracted and encoded from the inherent attribute fields of each entity node, denoted as f_static. For process chamber nodes, the extracted feature fields include chamber equipment model code (label code mapped to integers and normalized to [0,1]), chamber rated power (normalized based on the maximum rated value), chamber rated pressure (normalized using the same method), chamber cumulative operating days (normalized after taking the natural logarithm), and the production line number code to which the chamber belongs (unique thermal code). For component nodes, the component function type code, Boolean identifier of whether it is a shared component, and component rated life (normalized after taking the natural logarithm in hours) are extracted. For fault event nodes, the fault event level code (low, medium, and high severity are mapped to 0, 1, and 2 respectively and normalized) and fault event type code (unique thermal code) are extracted. For maintenance operation nodes, process recipe nodes, and production work order nodes, their attribute fields are extracted and encoded respectively. The static attribute feature vectors of all entity nodes are uniformly mapped to the same dimension d_s, and in the example, d_s=32.

[0048] Extending the dynamic operating feature matrix: First, extract the raw time-series data of the process chamber nodes within the observation window from the multi-dimensional time-series data of process parameters. Organize this data into a dynamic operating feature matrix M_dyn_raw, with the number of rows being the number of time steps T and the number of columns being the number of sensor channels C. In the example, T is 96 (corresponding to a 4-hour observation window, with a sampling step size of 2.5 minutes), and C is 32, resulting in a matrix dimension of 96×32. Next, calculate the moving mean, moving standard deviation, and moving slope sequences for each column of sensor channel data over T time steps. Finally, concatenate the dynamic operating feature matrix with the above three sets of moving statistics along the column direction to form an extended dynamic operating feature matrix M_dyn_ext, with dimensions T×(C×4), which is 96×128 in the example above.

[0049] The topological feature vector is extracted from the adjacency matrix of the dynamic heterogeneous graph. For each node, four graph topology metrics are calculated: node degree, node betweenness centrality, clustering coefficient, and node PageRank value. These metrics are then concatenated into a topological feature vector f_topo with a dimension of 4. For isolated nodes, all of the above metrics are set to 0.

[0050] (2) The temporal self-attention encoding channel is used to extract the runtime timing dependency features of the process chamber nodes from the extended dynamic runtime feature matrix and output the temporal encoding matrix H_time. After adding sinusoidal position encoding to the M_dyn_ext input position encoding layer, it is input to the stacked multi-head self-attention transformation module. Each module contains a multi-head self-attention sub-layer and a feedforward fully connected sub-layer, both of which use residual connections and layer normalization. The stacked layer is 3 layers in the example, the number of attention heads in each layer is 4 in the example, and the hidden layer dimension of each attention head is 32. After processing, the temporal encoding matrix H_time is obtained, with a dimension of T×d_model. In the example, d_model=128, and each row corresponds to the dependency feature vector of a time step.

[0051] (3) Graph attention encoding channel, used to extract topological neighborhood structure features between entity nodes from dynamic heterogeneous association graph, is implemented using a multi-relationship graph encoder based on graph attention network. The static attribute feature vector f_static of each entity node is concatenated with the topological structure feature vector f_topo as the initial node representation, which is processed through stacked multi-layer graph attention layers. Different types of associations are distinguished when calculating attention coefficients: a unified attention calculation mechanism is used for composition associations and component sharing associations; for fault attribution associations and fault cascading associations, an additional learnable association type embedding vector is introduced in the attention coefficient calculation to enable the model to distinguish the different influence strengths of different types of associations on node representations; for maintenance response associations and process adaptation associations, the corresponding type of learnable vector is embedded. The stacked layer number is 2 in the example, and the number of attention heads in each layer is 4 in the example. After processing, the structure encoding vector H_struct is obtained, with a dimension of d_gat, and d_gat=64 in the example.

[0052] (4) Cross-attention fusion layer, used to fuse the temporal encoding matrix H_time and the structural encoding vector H_struct across modalities to generate the spatiotemporal fusion feature vector H_fused. Using H_struct as the query vector and each time step vector of H_time as the key vector and value vector, the attention weights are calculated, normalized by softmax, and then the value vector matrix is ​​weighted and summed to obtain the structure-aware temporal context vector; this is concatenated with the original H_struct and compressed through a fully connected layer to obtain the spatiotemporal fusion feature vector H_fused, with a dimension of 128 in the example.

[0053] (5) Dual-task prediction output layer, used to receive H_fused, and input the parallel cumulative risk function estimation branch and the fault risk ranking branch respectively.

[0054] The cumulative risk function estimation branch employs discretized cumulative risk estimation. H_fused is input into a multilayer perceptron, outputting a K-dimensional vector, where K is the preset number of discrete time points. After sigmoid activation, each dimension represents the probability that the process chamber node is still in normal operation at that time point. In the example, K=20, corresponding to an evaluation point every hour from 1 to 20 hours in the future. To ensure the survival probability curve is monotonically non-increasing, the K-dimensional initial probability sequence output by the sigmoid is processed using an adjacent violation point pooling algorithm for order-preserving regression, and the survival probability at time 0 is forcibly set to 1. The training loss for this branch uses a negative log-likelihood loss function, employing different likelihood terms for truncated samples with fault events and censored samples without fault events.

[0055] Find the time point from the survival probability curve where the survival probability first drops to a preset probability threshold, and use the number of hours corresponding to that time point as the estimated remaining reliable runtime; the preset probability threshold is 0.5 in this example. If the survival probability does not drop below the preset probability threshold within K time points, then take the maximum time value among the K time points as the estimated remaining reliable runtime.

[0056] After outputting the estimated remaining reliable runtime, the maintenance process time is verified: combining the vacuum recovery time and clean recovery time of the current chamber corresponding to the maintenance level, as well as the sum of the total wafer loading and unloading time and the pure maintenance operation time calculated based on the wafer transfer speed and the current number of wafers being manufactured, the total maintenance process time is calculated; the remaining reliable runtime is compared with the total process time to mark whether the chamber currently has a sufficient safe intervention time window.

[0057] The fault risk ranking branch inputs H_fused into another multilayer perceptron, outputting an M-dimensional vector, where M is the total number of fault event types (in the example, M=10). After sigmoid activation, each dimension represents the independent probability of the corresponding type of fault event occurring in the future. The M probability values ​​are sorted from high to low, and the component with the highest probability is taken as the fault event risk score. The type corresponding to this component is then taken as the most probable fault event type.

[0058] The training process of the joint prediction framework includes: training sample construction, for each process chamber node, taking the start time of the observation window as the benchmark, extracting multi-dimensional time-series data of process parameters within the observation window and a snapshot of the dynamic heterogeneous correlation graph at the start time of the window as input, and using the actual health status within a preset evaluation period after the end time of the observation window as the label; the joint loss function is a weighted sum of the negative log-likelihood loss of the cumulative risk function estimation branch and the binary cross-entropy loss of the fault risk ranking branch, with the weight of the loss of both branches set to 0.5 in the example; parameter optimization uses the AdamW optimizer with an initial learning rate of 0.0005, a weight decay coefficient of 0.01, and a cosine annealing learning rate scheduling strategy, with a maximum training epoch of 200 epochs and a batch size of 16 process chamber nodes in the example.

[0059] S400. Construct a hierarchical multi-agent collaborative decision-making framework and output maintenance scheduling decisions: Construct a hierarchical multi-agent collaborative decision-making framework, mapping each process chamber to a chamber maintenance agent, and setting a global maintenance arbitration agent; Before generating a maintenance request, each chamber maintenance agent performs physical constraint verification based on the vacuum recovery time, clean recovery time, remaining capacity of the matching wafer buffer cavity, and minimum continuous processing batch constraint of the process chamber node, and performs maintenance window feasibility prediction based on the estimated remaining reliable runtime. When any physical constraint is not met, the generation of preventive maintenance requests and emergency maintenance requests is prohibited; Each chamber maintenance agent generates maintenance requests based on its own local observation state and the shared state information of adjacent chamber maintenance agents. The global maintenance arbitration agent resolves conflicts based on the maintenance requests of all chamber maintenance agents and the global production state, and outputs a maintenance scheduling decision for arranging semiconductor equipment chamber maintenance operations.

[0060] Map each process chamber to a chamber maintenance agent and set up a global maintenance arbitration agent, as shown in the example below. Figure 4 As shown.

[0061] The chamber maintenance agent is responsible for generating and executing maintenance requests for the corresponding process chamber node and providing status feedback. Its state space includes the following five state components: The first state component is the estimated remaining reliable runtime of the process chamber node output by the joint prediction framework, in hours; the second state component is the type of failure event with the highest probability of occurrence for the process chamber node and its risk score, which is in the range [0,1]; the third state component is the completion progress of the production work order currently undertaken by the process chamber node, with a value of 0 to 1; the fourth state component is the planned estimated maintenance time for the process chamber node; and the fifth state component is the estimated maintenance time of other chamber maintenance agents that have edge connections with the process chamber node in the collaborative communication graph.

[0062] The action space of the chamber maintenance agent includes: Action 1 is not initiating a maintenance request, used in scenarios where the estimated remaining reliable runtime is greater than a preset safety threshold and the fault event risk score is lower than a preset risk threshold; Action 2 is initiating a preventive maintenance request, used in scenarios where the estimated remaining reliable runtime is less than or equal to a preset safety threshold; Action 3 is initiating an emergency maintenance request, used in scenarios where the fault event risk score is greater than or equal to a preset risk threshold, with the preset risk threshold set to 0.8 in an example; Action 4 is requesting delayed maintenance, used in scenarios where the delivery deadline of the currently accepted production work order is less than a preset delivery deadline threshold, the task priority is higher than a preset priority, and the estimated remaining reliable runtime is within a set range.

[0063] Before selecting an action, the chamber maintenance agent must first complete physical constraint verification. If any one of these conditions is not met, the agent cannot select to initiate preventive maintenance requests or emergency maintenance requests. The agent can only select not to initiate a maintenance request or request delayed maintenance: wafer capacity verification, i.e., the remaining capacity of the matching Loadlock is greater than or equal to the total number of wafers currently being manufactured in the chamber; batch continuity verification, i.e., the current batch has been completed, or the remaining processing volume meets the minimum continuous processing batch requirement; window duration verification, i.e., the current available window duration is greater than or equal to the total time consumed in the entire maintenance process.

[0064] Before generating a maintenance request, the chamber maintenance agent performs a maintenance window feasibility prediction, which includes: (1) extracting the process path of the production work order currently accepted by the process chamber node and identifying the key process nodes. The key process nodes refer to the processing steps with strict time window constraints in the process path, such as the development process must be completed within 4 hours after the photolithography process and the annealing process must be completed within 8 hours after the ion implantation process; (2) calculating the constraint interval of the maintenance available window. The starting point of the window is the current time, and the ending point of the window is the planned start time of the next key process node minus the process preparation time, and then minus the pure maintenance operation time, vacuum recovery time, clean recovery time, and total wafer loading and unloading time; (3) when the estimated value of the remaining reliable runtime exceeds the constraint interval of the above maintenance available window, that is, the remaining reliable runtime is insufficient to support the completion of the current production work order, the chamber maintenance agent automatically generates a process path reordering request and applies to the global maintenance arbitration agent to adjust the non-critical process tasks in the current production work order of the process chamber node to be executed before the critical process tasks, thereby reserving a feasible execution time window for maintenance operations.

[0065] The global maintenance arbitration agent is responsible for coordinating and arbitrating global maintenance requests. Its state space includes the following five state components: The first state component is a list of estimated maintenance times for all chamber maintenance agents, where each item contains the process chamber node number and the corresponding estimated maintenance start timestamp; the second state component is the completion progress of the overall semiconductor processing scheduling tasks of the equipment cluster, measured by the equipment cluster-level task completion rate, calculated as the total number of substrates processed by all process chamber nodes in the equipment cluster at the current moment divided by the planned total number of wafers processed by all production orders in the equipment cluster; the third state component is the task priority within the equipment cluster. The proportion of production work orders with a priority higher than the preset priority is calculated by dividing the sum of the planned processing number of high-priority production work orders by the sum of the planned processing number of all production work orders. High-priority production work orders refer to production work orders whose delivery deadline is less than the preset time threshold or whose priority is marked as the highest level. The fourth state component is the current health status of the components in the shared list of components within the equipment cluster, which is measured by the ratio of the cumulative running time of each component to its rated life, and the value is in the range of [0,1]. The fifth state component is the current occupancy status of the components in the shared list of components within the equipment cluster, which is represented by a Boolean value.

[0066] The action space of the global maintenance arbitration agent includes: Action 1 is to approve maintenance requests, which is used for scenarios where the requests meet the constraints and will not cause resource conflicts; Action 2 is to reject maintenance requests, which is used for scenarios where the requests violate the constraints and the conflicts cannot be eliminated by adjustment. When rejecting a request results in an unplanned downtime risk for the device exceeding a preset threshold, the global maintenance arbitration agent will forcibly approve the maintenance request and trigger a load transfer strategy; Action 3 is to adjust the maintenance time, which is used for scenarios where the requests violate the constraints but other feasible time windows exist.

[0067] After receiving a process path reordering request, the global maintenance arbitration agent assesses the impact of adjusting the non-critical process task in the current production work order of the process chamber node to be executed before the critical process task on other process chamber nodes in the equipment cluster, and counts the number of process chamber nodes affected by the path reordering. If the number of affected process chamber nodes is less than or equal to a preset threshold (2 in this example), and will not cause production work orders with higher priority than the preset priority to be overdue, the rescheduling request is approved, and the maintenance schedule of the relevant process chamber nodes is adjusted globally. If the number of affected process chamber nodes is greater than the preset threshold, the impact of the rescheduling on high-priority production work orders is first assessed: if it will not cause high-priority work orders to be overdue, a partial rescheduling is approved, adjusting only the process sequence of low-priority work orders, while transferring the affected non-critical processes to a spare process chamber for execution; if it will cause high-priority work orders to be overdue, the rescheduling request is rejected, and the nearest mandatory maintenance window is planned for the chamber maintenance agent that initiated the request, and a load transfer instruction is sent to adjacent healthy chambers to gradually transfer the work orders in progress in that chamber to idle chambers, completing maintenance before a failure occurs and avoiding unplanned downtime.

[0068] Before approving a maintenance request, the global maintenance arbitration agent performs a global capacity balance assessment, calculating the impact coefficient of the maintenance decision on the overall capacity of the equipment cluster. The capacity impact coefficient is calculated as follows: the number of process chamber nodes that will be shut down due to maintenance divided by the total number of process chamber nodes in the equipment cluster, multiplied by the estimated maintenance time divided by the standard production cycle. When the capacity impact coefficient exceeds a preset critical threshold (0.3 in the example), the global maintenance arbitration agent activates a capacity protection mode: Strategy 1: Prioritize the approval of emergency maintenance requests with an estimated remaining reliable runtime below the emergency threshold (24 hours in the example); Strategy 2: For preventative maintenance requests with an estimated remaining reliable runtime above the lenient threshold (48 hours in the example), force the execution time to be delayed to a period of lower production load; Strategy 3: For process chamber nodes whose estimated remaining reliable runtime is less than the safety threshold and must be maintained immediately, a load transfer instruction is sent to other healthy chamber maintenance agents through the collaborative communication graph, temporarily transferring their current production work order to a standby process chamber node with idle capacity.

[0069] When maintenance requests submitted by multiple chamber maintenance agents overlap in time windows and the overlapping maintenance tasks involve the same shared components, the global maintenance arbitration agent performs component collaborative maintenance optimization, specifically including: (1) identifying the aggregation degree of maintenance requirements for shared components. The aggregation degree is calculated by multiplying the normalized value of the wear degree of the component and the normalized value of the fault event risk score of each process chamber node sharing the component. The aggregation degree value is obtained by summing the product of the components of all process chamber nodes and dividing it by the total number of process chamber nodes sharing the component. The normalized value of the wear degree of the component is obtained by dividing the cumulative running time of the component by its rated life and truncating it to the [0,1] interval. The normalized value of the fault event risk score is obtained directly from the fault event risk score output by the joint prediction framework. (2) When the aggregation degree value exceeds the preset aggregation degree threshold (0.7 in the example), the global maintenance arbitration agent will merge the maintenance requests of multiple process chamber nodes involving the shared component into a batch maintenance task, uniformly allocate maintenance resources, and complete the replacement or maintenance of the shared component in one go. (3) When the aggregation degree value is less than or equal to the preset aggregation degree threshold, the global maintenance arbitration agent will approve each maintenance request in the order of arrival time of the request. Furthermore, after the maintenance task of the shared component is completed, the subsequent chamber maintenance agent will delay its own maintenance execution time by at least 4 hours to avoid secondary wear of the component caused by frequent maintenance.

[0070] The reward function for the global maintenance arbitration agent is defined as follows: the total reward equals the cumulative output reward plus the collaborative reward, minus the maintenance cost penalty, and then minus the penalty for sudden failure events. The cumulative output reward is calculated by multiplying the number of substrates processed in the current production order undertaken by the process chamber node by the output value weight per unit substrate, and summing this sum across all online process chamber nodes. The collaborative reward is calculated by: for the process chamber node and each of its adjacent process chamber nodes with edge connections in the collaborative communication graph, determining whether the time interval between their expected maintenance times meets the preset collaborative constraints. If the time interval is greater than or equal to the preset minimum collaborative interval (4 hours in the example), the collaborative interval compliance flag is set to 1; otherwise, it is set to 0. The collaborative reward is obtained by multiplying the collaborative interval compliance flag by the collaborative reward adjustment coefficient and summing this sum across all adjacent node pairs. The maintenance cost penalty is calculated as the maintenance labor cost plus the unit cost of replacing components multiplied by the actual number of components replaced. The calculation method for the penalty for sudden failure events is to multiply the penalty base coefficient for the failure event by the level coefficient of the sudden failure event occurring at the process chamber node, then multiply by the actual downtime of the equipment caused by the failure event, and finally multiply by the hourly capacity value of the equipment cluster.

[0071] The construction of the collaborative communication graph includes: (1) selecting process chamber nodes from all entity nodes in the dynamic heterogeneous association graph as nodes in the collaborative communication graph, with each node corresponding to a chamber maintenance agent. (2) extracting the basic attributes and dynamic collaborative attributes of each process chamber node from the dynamic heterogeneous association graph as node attributes. The basic attributes include the process chamber number, process chamber equipment model, and rated operating parameters of the process chamber. The dynamic collaborative attributes include the estimated maintenance time and the current health status. The estimated maintenance time is extracted by finding a triplet in the dynamic heterogeneous association graph with the process chamber node as the subject entity node, the maintenance response association as the association relationship, and the maintenance operation node as the object entity node. The maintenance operation execution time is extracted from its time sequence attributes. If there are multiple triplets, the latest execution time is selected. The current health status is deduced by searching for a triplet in the dynamic heterogeneous association graph with the process chamber node as the subject entity node and the fault attribution association as the association relationship. If it does not exist in the most recent preset observation period (the past 72 hours in the example), it is marked as normal. If it exists, the fault event level and occurrence time are extracted and marked as the corresponding fault event level. (3) The edges of the collaborative communication graph are generated based on triplets in the dynamic heterogeneous association graph where both the subject and object entity nodes are process chamber nodes: undirected edges are established between the corresponding chamber maintenance agents based on the component sharing association triplet; directed edges are established between the corresponding chamber maintenance agents based on the fault cascade association triplet, with the direction from the fault event source node to the fault event potential impact node.

[0072] See Figure 5 This embodiment provides a feasible method for output maintenance scheduling decision-making, including steps S410 to S480: S410. Each chamber maintenance agent, based on its current state, outputs the probability distribution of each action in the action space through its Actor network, and selects an action according to the ε-greedy strategy. If the selected action belongs to the type that needs to submit a request to the global maintenance arbitration agent, a maintenance request containing the request type, estimated time consumption, and the latest acceptable start time is generated.

[0073] S420. Each chamber maintenance agent synchronizes its estimated maintenance time to the adjacent chamber maintenance agents through the cooperative communication graph. At the same time, it obtains the latest estimated maintenance time of the adjacent chamber maintenance agents from the cooperative communication graph and updates the fifth state component in its own state space.

[0074] S430: The global maintenance arbitration agent collects all maintenance requests and their current states. Based on its own state space, it outputs the selection probability distribution of each coordinated action through its Actor network and selects actions according to the ε-greedy strategy. Before selecting an action, the global maintenance arbitration agent sequentially executes three mechanisms: maintenance window feasibility prediction, global capacity balance assessment, and component collaborative maintenance optimization. The processing results of each mechanism are used as constraints to introduce the action selection process.

[0075] S440. Based on the action execution results of the global maintenance arbitration agent, calculate the total reward according to the reward function, and synchronously feed back the total reward to each chamber maintenance agent that submitted a maintenance request.

[0076] S450: Record the state and selected actions of all chamber maintenance agents, the state and selected actions of the global maintenance arbitration agent, and the corresponding state and actions at the next moment, along with the edge connections of the collaborative communication graph in this decision, and store them in the experience replay pool.

[0077] S460. Randomly sample a batch of experience data from the experience replay pool. Sequentially concatenate the state vectors and action vectors of all chamber maintenance agents and global maintenance arbitration agents within the batch into a global state-action joint vector. Input the vector into the main Critic network to calculate the global state-action value estimate at the current moment. Calculate the Critic network loss based on the temporal difference error between the estimate and the target global state-action value estimate at the next moment. Use the Adam optimizer to minimize the loss to update the main Critic network parameters and update the target Critic network parameters according to the soft update strategy.

[0078] S470. Fix the updated main Critic network parameters, concatenate the current actions generated by the Actor networks of each chamber maintenance agent in the current state and the current actions generated by the Actor network of the global maintenance arbitration agent in the current state, and input them into the main Critic network to obtain the global state-action value estimate under the current policy. Calculate the policy gradient loss of each Actor network separately with the optimization objective of maximizing this value. Use the Adam optimizer to update the Actor network parameters of all chamber maintenance agents and the global maintenance arbitration agent respectively, and update the corresponding target Actor network parameters respectively according to the soft update policy.

[0079] S480. After training convergence, only the Actor networks of each chamber maintenance agent and the global maintenance arbitration agent are retained for online inference. During online inference, the Actor network of the chamber maintenance agent receives the real-time state vector of the corresponding process chamber node to generate maintenance request actions, and the Actor network of the global maintenance arbitration agent receives the global real-time state vector to generate coordination decision actions.

[0080] Example 2: Maintenance of a 12-inch Plasma Etching Equipment Cluster This embodiment uses a cluster-type plasma etching equipment cluster of a 12-inch integrated circuit production line as the application object. The equipment includes four process chambers: CH-A01, CH-A03, CH-A05, and CH-A07, all belonging to Class 100 cleanroom zone Z1 and sharing the same vacuum main circuit group V-G01. Among them, CH-A03 and CH-A05 share a high-frequency radio frequency power supply of model RF-800W (component number CP-205). The current system time is 10:00:00 on June 20, 2024, and the equipment cluster is executing three batches of logic chip etching work orders.

[0081] Preset physical constraint benchmark parameters for semiconductor equipment: Vacuum system constraints: Standard time for evacuating to process vacuum (5mTorr) after single-chamber vacuum breaking is 60 minutes; vacuum leak detection verification time is 15 minutes; the vacuum main circuit group V-G01 can support a maximum of 2 chambers for parallel vacuuming simultaneously; Cleanliness constraints: Total time for nitrogen purging and particle concentration verification after open-cover maintenance is 30 minutes; the maximum number of chambers allowed to be opened simultaneously in clean zone Z1 is 2; Transfer and process constraints: Standard time for single-step wafer transfer is 0.8 minutes per wafer; single-chamber rated wafer carrier is 25 wafers per batch; the total capacity of the matching Loadlock buffer chamber is 50 wafers, of which 10 are currently occupied; the minimum continuous processing batch size for a single batch of wafers is 10 wafers; the hard constraint for etching Q-Time after photolithography is 8 hours; Maintenance operation benchmarks: Preventive open-cover maintenance (replacing RF power supply) single-chamber maintenance operation time is 120 minutes; preventive cleaning maintenance operation without opening the cover takes 60 minutes.

[0082] Full operational and production data from the past 4 hours (06:00-10:00) was collected and standardized to obtain a standardized multi-source dataset. Key values ​​are as follows: Multi-dimensional time-series data of process parameters, comprising 32 sensor channels, with a sampling step size of 2.5 minutes and a total of 96 time steps. Specifically, the average RF reflection power window value for CH-A03 is 12.5W, with a standardized Z-score of 0.18; the average chamber pressure window value is 4.8 mTorr, with a standardized value of -0.07; and the wafer transfer speed is consistently 0.8 minutes per wafer. The average RF reflection power window value for CH-A05 is 11.8W, with a standardized value of 0.14. Equipment maintenance history data: the last preventative maintenance for CH-A03 was on May 20, 2024, corresponding to 850 hours of cumulative CP-205 operation. The last maintenance for CH-A05... The date is May 22, 2024, corresponding to 820 hours of cumulative operation for CP-205. The rated lifespan of the CP-205 RF power supply is 1000 hours. Semiconductor processing scheduling data: CH-A03 received work order LOT-5621, planned to process 50 wafers, high priority, delivery deadline 24:00 on the same day, current completion progress 60% (30 wafers processed, 15 wafers manufactured in the cavity), next critical etching step planned start time is 22:00 on the same day; CH-A05 received work order LOT-5623, planned to process 50 wafers, high priority, current completion progress 52% (12 wafers manufactured in the cavity), next critical etching step planned start time is 23:00 on the same day; Supplementary data for equipment static attributes includes the above-mentioned vacuum, cleanliness, and transmission constraint parameters, which, together with the original chamber rated parameters and component ledger, complete the attribute filling.

[0083] A dynamic heterogeneous association graph was constructed based on a standardized dataset: 26 entity nodes across 6 categories were defined, including 4 process chamber nodes, 12 component nodes, 3 fault event nodes, 4 maintenance operation nodes, 2 process recipe nodes, and 3 production work order nodes; 37 association edges were established across 6 categories, including 12 component associations, 2 component sharing associations, 1 fault cascading association, 6 process adaptation associations, 8 maintenance response associations, and 8 fault attribution associations. Based on the component sharing association between CH-A03 and CH-A05, the fault cascading association was automatically derived and the propagation intensity coefficient was calculated: The time difference between the last maintenance of the two chambers was ΔT = 48 hours, and the attenuation rate parameter α was set to 0.2. Therefore, the propagation intensity coefficient = 1 / (1+e^(-0.2×48))≈0.9999, indicating a very high risk of fault propagation due to the shared RF power supply between the two chambers. After construction, a graph index was established for chamber numbers and association time attributes, with a single-step graph query response latency of less than 10ms.

[0084] The dynamic heterogeneous correlation graph was input into the joint prediction framework. The future prediction period was set to 24 hours, with an evaluation point every hour (K=24), and the survival probability judgment threshold was 0.5. The prediction results for each process chamber were output as shown in Table 1.

[0085] Table 1. Prediction results for each process chamber

[0086] After outputting the prediction results, a full maintenance process time verification was performed. Taking the CH-A03 RF power supply replacement after opening the cover as an example, the total downtime of the entire maintenance process is equal to the maintenance operation time of 120 minutes, the total vacuum recovery time of 75 minutes, the clean recovery time of 30 minutes, and the total wafer loading and unloading time (15 wafers × 0.8 minutes per wafer × 2 = 24 minutes), totaling 249 minutes, approximately 4.15 hours. After verification, the remaining reliable running time of CH-A03 is 10.5 hours, which is greater than the total downtime of the entire process, thus meeting the basic conditions for safe intervention.

[0087] Each chamber maintenance agent makes decisions based on its own state and information shared with neighboring agents. Taking CH-A03 as an example: its state space has a remaining reliable runtime of 10.5 hours, the most probable fault type is radio frequency power drift anomaly, the fault risk score is 0.85, the current production work order completion progress is 0.6, it has no planned estimated maintenance time, and the neighboring chamber CH-A05 in the collaborative communication diagram has no estimated maintenance time. Physical constraint pre-verification: Batch continuity verification: The current batch has 10 wafers remaining for continuous processing, meeting the minimum continuous processing batch requirement, thus compliant; Wafer capacity verification: 15 wafers are manufactured inside the CH-A03 cavity, and the remaining available capacity of the matching Loadlock is 40 wafers, thus compliant; Window duration verification: Entering the maintenance window feasibility prediction process, the window start time is the current 10:00, and the window end time is equal to the start time of the next critical process 22:00 minus the process preparation time of 1 hour minus the total downtime of the entire maintenance process of 4.15 hours, which is 16:51 on the same day. The remaining reliable running time can support until 20:30, which is later than the window end time, so the window is feasible. Action selection and request generation: The fault risk score of 0.85 is greater than or equal to the preset emergency threshold of 0.8, therefore an emergency maintenance request is initiated. The maintenance type is opening the cover to replace the RF power supply, and the acceptable latest start time is 16:51 on the same day. The request information is simultaneously submitted to the global maintenance arbitration intelligent agent. Similarly, after verification, the CH-A05 intelligent agent selects to initiate a preventive maintenance request, with the latest acceptable start time being 17:37 on the same day.

[0088] After receiving all requests, the global maintenance arbitration agent sequentially performs global resource constraint verification, conflict resolution, and component collaborative optimization based on its own global state space. The list of shared components in the global state space includes the matching Loadlock buffer cavity, currently with a total capacity of 50 wafers, 10 wafers occupied, and 40 wafers remaining. The cleanroom zone Z1 currently has 0 open-cap operations, and the vacuum pipeline group V-G01 currently has 0 vacuuming operations. Global resource constraint verification: Wafer capacity verification: The CH-A03 and CH-A05 cavities contain a total of 27 wafers, which can be fully accommodated by the 40 wafers remaining in the matching Loadlock, thus complying with regulations; Cleanroom zone verification: The number of simultaneously open-cap chambers is 2, which does not exceed the maximum quota of 2 in zone Z1, thus complying with regulations; Vacuum pipeline verification: The two chambers are simultaneously vacuumed, which does not exceed the parallel limit of 2 in group V-G01, thus complying with regulations.

[0089] Component Collaborative Maintenance Optimization: The aggregation degree of maintenance requirements for the shared component CP-205 is calculated. The normalized wear value for component CH-A03 is 850 / 1000 = 0.85, the fault risk score is 0.85, and the product is 0.7225. The normalized wear value for component CH-A05 is 820 / 1000 = 0.82, the fault risk score is 0.82, and the product is 0.6724. The aggregation degree = (0.7225 + 0.6724) / 2 ≈ 0.697. Components exceeding the preset threshold of 0.7 are judged to undergo batch merged maintenance after rounding the aggregation degree or adjusting for collaborative benefits. As an optional implementation, the aggregation degree can be adjusted based on the collaborative benefits of physical constraints. The combined maintenance of the two chambers can share the vacuuming and clean purging processes, saving approximately 4.2 hours of downtime. After normalizing the collaborative benefits, the aggregation degree is adjusted to 0.743, exceeding the preset threshold of 0.7, therefore batch merged maintenance is determined. The combined batch maintenance operation took 150 minutes. The vacuum recovery and cleanroom recovery processes were executed together, with a total recovery time of 75 minutes. Wafer loading and unloading were carried out simultaneously, with a total time of 24 minutes. The total downtime for the entire maintenance process was 249 minutes, which is 249 minutes less downtime than maintaining the two chambers separately. The combined maintenance execution sequence was determined as follows: the operation started at 11:00 AM, with wafers being unloaded from both chambers simultaneously. At 11:24 AM, the cover was opened to replace the CP-205 RF power supply. Mechanical operations were completed at 1:54 PM, with vacuuming and cleanroom purging performed simultaneously. Vacuum leak detection and particle detection were completed at 3:09 PM, and wafer loading was completed at 3:33 PM, restoring the equipment to a process-usable state.

[0090] The total reward calculation for this decision is as follows: Cumulative output reward: avoiding production interruptions caused by unplanned failures and recovering 100 wafers of capacity from two batches, with a unit substrate output weight of 200 yuan per wafer, the cumulative output reward = 100 × 200 = 20,000 yuan; Collaboration reward: the combined maintenance of the two chambers saves 4.15 hours of downtime, with a collaboration reward adjustment coefficient of 500 yuan per hour, the collaboration reward = 4.15 × 500 = 2,075 yuan; Maintenance cost penalty: labor cost of 1,500 yuan and RF power supply component cost of 8,000 yuan, totaling 9,500 yuan; Total reward = 20,000 + 2,075 - 9,500 = 12,575 yuan.

[0091] The global maintenance arbitration agent outputs the final decision: approves the emergency maintenance request for CH-A03 and the preventive maintenance request for CH-A05, merging them into a batch maintenance task for common components, to be executed from 11:00 to 15:33 on June 20, 2024; CH-A01 and CH-A07 will remain in normal operation and will not be scheduled for maintenance for the time being, with the next status assessment cycle 4 hours later; the process paths for LOT-5621 and LOT-5623 will be fine-tuned, adjusting the non-critical pre-cleaning steps to be executed after maintenance, ensuring that the critical etching steps are completed within the Q-Time hard constraint and without affecting the work order delivery deadline.

[0092] The decision results of this embodiment are compared with the traditional periodic maintenance strategy and the single-cavity independent predictive maintenance strategy. The core indicators are shown in Table 2.

[0093] Table 2 Comparison of Core Indicators

[0094] This method, by embedding semiconductor-specific physical hard constraints and a global group decision-making mechanism, can significantly reduce the risk of unplanned downtime and total maintenance costs while ensuring the physical feasibility of maintenance solutions, and achieve global optimization matching between maintenance schedules and production tasks.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described with reference to preferred embodiments, those skilled in the art should understand that various changes in form and detail can be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A semiconductor device chamber maintenance group decision-making method, characterized in that, include: By collecting multi-dimensional time-series data of process parameters through sensors deployed in the semiconductor equipment chamber, and collecting equipment maintenance history data, process specification definition data, semiconductor processing scheduling data and equipment static attribute data, a standardized multi-source dataset is obtained through standardization. Based on the standardized multi-source dataset, six types of entity nodes are defined: process chamber nodes, component nodes, fault event nodes, maintenance operation nodes, process recipe nodes, and production work order nodes. According to six types of association relationships between entity nodes—composition association, fault attribution association, maintenance response association, process adaptation association, component sharing association, and fault cascading association—a dynamic heterogeneous association graph is constructed by combining the timestamp attribute, effective time interval attribute, and data sampling granularity attribute of each association relationship. Specifically, based on the component sharing association between two process chamber nodes, the fault cascading association between the two process chamber nodes is automatically derived, and the fault propagation risk is characterized by a fault propagation intensity coefficient, which increases as the time difference between the most recent maintenance operation of the two process chamber nodes increases. Based on the dynamic heterogeneous correlation graph, a joint prediction framework of dual-channel graph temporal coding and cumulative risk function estimation is adopted. Dynamic runtime sequence features and topological neighborhood structure features are extracted for each process chamber node. After fusion through a cross-attention mechanism, the survival probability curve of each process chamber node in the future preset time period is output. The time point corresponding to the first drop in survival probability to the preset probability threshold is found from the survival probability curve. The duration corresponding to the time point is used as the estimated value of the remaining reliable runtime of the process chamber node. The fault event type with the highest probability of occurrence and its risk score for each process chamber node are output. A hierarchical multi-agent collaborative decision-making framework is constructed, mapping each process chamber to a chamber maintenance agent and setting a global maintenance arbitration agent. Before generating a maintenance request, each chamber maintenance agent performs physical constraint verification based on the vacuum recovery time, clean recovery time, remaining capacity of the associated wafer buffer cavity, and minimum continuous processing batch constraint of the process chamber node. It also performs maintenance window feasibility prediction based on the estimated remaining reliable runtime. If any physical constraint is not met, the generation of preventive maintenance requests and emergency maintenance requests is prohibited. Each chamber maintenance agent generates maintenance requests based on its own local observation state and the shared state information of adjacent chamber maintenance agents. The global maintenance arbitration agent resolves conflicts based on the maintenance requests of all chamber maintenance agents and the global production state, and outputs a maintenance scheduling decision for arranging semiconductor equipment chamber maintenance operations.

2. The semiconductor device chamber maintenance group decision-making method according to claim 1, characterized in that, The construction of the dynamic heterogeneous association graph includes: Triples are extracted from the multi-source data of the standardized multi-source dataset. The triple format is subject entity node, association relationship, object entity node, and time sequence attribute. Create entity nodes and populate attributes for the extracted triples; Establish associations and bind temporal attributes to the extracted triples; The extracted triples are used to derive and supplement fault cascading associations, including: when process chamber node X and process chamber node Y have a component-shared association through component node Z, a fault cascading association is automatically established between process chamber node X and process chamber node Y, and the fault propagation intensity is set according to the fault propagation intensity coefficient.

3. The semiconductor device chamber maintenance group decision-making method according to claim 1, characterized in that, The dual-channel graph temporal coding and cumulative risk function estimation joint prediction framework includes a multi-dimensional feature construction layer, a temporal self-attention coding channel, a graph attention coding channel, a cross-attention fusion layer, and a dual-task prediction output layer; The multidimensional feature construction layer is used to construct static attribute feature vectors, expand dynamic running feature matrices, and topological structure feature vectors for each entity node based on the dynamic heterogeneous association graph. The temporal self-attention coding channel is used to extract the runtime timing dependency features of the process chamber nodes from the extended dynamic runtime feature matrix and output the temporal coding matrix; The graph attention encoding channel is used to extract the topological neighborhood structure features between entity nodes from the dynamic heterogeneous association graph, output the structure encoding vector, and introduce a learnable association type embedding vector for fault attribution association and fault cascade association when calculating the attention coefficient, so as to distinguish the influence intensity of different types of association relationships on node representation. The cross-attention fusion layer is used to perform cross-modal fusion of the temporal coding matrix and the structural coding vector to generate a spatiotemporal fusion feature vector for each process chamber node. The dual-task prediction output layer is used to receive the spatiotemporal fusion feature vector and input the parallel cumulative risk function estimation branch and the fault risk ranking branch respectively. The cumulative risk function estimation branch is used to output the survival probability curve of each process chamber node in the future preset time period, and to perform order-preserving regression post-processing on the survival probability curve to make it monotonically non-increasing. The fault risk ranking branch is used to output the fault event type with the highest probability of occurrence for each process chamber node and its risk score.

4. The semiconductor device chamber maintenance group decision-making method according to claim 1, characterized in that, The state space of the chamber maintenance agent includes: the estimated remaining reliable runtime of the process chamber node output by the joint prediction framework, the type of failure event with the highest probability of occurrence of the process chamber node output by the joint prediction framework and its risk score, the completion progress of the production work orders currently undertaken by the process chamber node, the planned maintenance time of the process chamber node, and the estimated maintenance time of other chamber maintenance agents that have edge connections with the process chamber node in the collaborative communication graph. The action space of the chamber maintenance agent includes not initiating a maintenance request, initiating a preventive maintenance request, initiating an emergency maintenance request, and requesting delayed maintenance; The state space of the global maintenance arbitration agent includes a list of estimated maintenance times for all chamber maintenance agents, the completion progress of the overall semiconductor processing scheduling tasks of the equipment cluster, the proportion of production work orders with task priorities higher than preset priorities within the equipment cluster, and the current health status and current occupancy status of components in the shared list of components within the equipment cluster. The action space of the global maintenance arbitration agent includes approving maintenance requests, rejecting maintenance requests, and adjusting maintenance time. The reward function of the global maintenance arbitration agent is defined as the total reward, which is the cumulative output reward plus the collaborative reward, minus the maintenance cost penalty, and then minus the penalty for sudden failure events.

5. The semiconductor device chamber maintenance group decision-making method according to claim 4, characterized in that, The feasibility assessment of the maintenance window includes: Extract the process path of the production work order currently accepted by the process chamber node, and identify the key process nodes with time window constraints; The constraint interval of the maintenance available window is calculated. The window starts at the current time and ends at the planned start time of the next critical process node minus the process preparation time, and then minus the pure maintenance operation time, vacuum recovery time, clean recovery time, and total wafer loading and unloading time. When the estimated remaining reliable runtime exceeds the constraint range of the maintenance available window, the chamber maintenance agent automatically generates a process path reordering request and applies to the global maintenance arbitration agent to adjust the non-critical process tasks in the current production work order of the process chamber node to be executed before the critical process tasks.

6. The semiconductor device chamber maintenance group decision-making method according to claim 5, characterized in that, After receiving a process path reordering request, the global maintenance arbitration agent assesses the impact of the task adjustment on other process chamber nodes in the equipment cluster and counts the number of process chamber nodes affected by the path reordering. If the number of affected process chamber nodes is less than or equal to the preset threshold, and will not cause the delivery of production work orders with a task priority higher than the preset priority to be overdue, then the reordering request will be approved, and the maintenance schedule of the relevant process chamber nodes will be adjusted globally. If the number of affected process chamber nodes exceeds the preset threshold, the impact of reordering on production work orders with task priorities higher than the preset priority is further evaluated: if it will not cause them to be late for delivery, a partial reordering is approved, only adjusting the process sequence of production work orders with task priorities lower than the preset priority, and transferring the affected non-critical processes to a spare process chamber for execution; if it will cause them to be late for delivery, the reordering request is rejected, a forced maintenance window is planned for the chamber maintenance agent that initiated the request, and a load transfer instruction is sent to the adjacent chamber maintenance agents that are in a healthy state to transfer the work orders of the process chamber node to an idle process chamber node for execution.

7. The semiconductor device chamber maintenance group decision-making method according to claim 4, characterized in that, When maintenance requests submitted by multiple chamber maintenance agents overlap in time windows, and the overlapping maintenance tasks involve the same shared components, the global maintenance arbitration agent performs component collaborative maintenance optimization, including: The aggregation degree of maintenance requirements for shared components is identified. The aggregation degree is calculated as follows: for each process chamber node sharing the component, the product of the normalized value of the component wear degree and the normalized value of the fault event risk score of the process chamber node is calculated. The product of all process chamber nodes is summed and divided by the total number of process chamber nodes sharing the component to obtain the aggregation degree value. The normalized value of the component wear degree is obtained by dividing the cumulative running time of the component by its rated life and truncating it to the [0,1] interval. When the aggregation degree value exceeds the preset aggregation degree threshold, the global maintenance arbitration agent will merge the maintenance requests of multiple process chamber nodes involving the common component into a batch maintenance task, and complete the replacement or maintenance of the common component in one go. When the aggregation degree value is less than or equal to the preset aggregation degree threshold, the global maintenance arbitration agent approves each maintenance request in the order in which the requests arrive.

8. The semiconductor device chamber maintenance group decision-making method according to claim 4, characterized in that, The output maintenance scheduling decision includes: Each chamber maintenance agent, based on its current state, outputs the probability distribution of each action in the action space through the Actor network of the chamber maintenance agent, and selects an action according to the ε-greedy strategy. If the selected action belongs to the type that needs to submit a request to the global maintenance arbitration agent, a maintenance request containing the request type, estimated time and the latest acceptable start time is generated. Each chamber maintenance agent synchronizes its estimated maintenance time to neighboring chamber maintenance agents through the cooperative communication graph, and obtains the latest estimated maintenance time of neighboring chamber maintenance agents from the cooperative communication graph to update its own state space. The global maintenance arbitration agent collects all maintenance requests and current states. Based on its own state space, it outputs the probability distribution of coordinated action selection through the Actor network of the global maintenance arbitration agent. It selects actions according to the ε-greedy policy, calculates the total reward according to the reward function, and feeds back the total reward to each chamber maintenance agent that submitted the request. The states and actions of all chamber maintenance agents and global maintenance arbitration agents at each time step, along with the edge connections of the cooperative communication graph, are stored in the experience replay pool. Samples are taken from the experience replay pool, and the state vectors and action vectors of each agent in the batch are concatenated into a global state-action joint vector, which is then input into the main Critic network. The parameters of the main Critic network are updated based on the temporal difference error, and the parameters of the target Critic network are softly updated. The parameters of the main Critic network are fixed, and the Actor network parameters of each chamber maintenance agent and global maintenance arbitration agent are updated respectively with the goal of maximizing the global state-action value estimate, and the corresponding target Actor network parameters are softly updated. After training convergence, only the Actor network of the maintenance agents for each chamber and the global maintenance arbitration agent is retained for online inference.

9. An electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it causes the electronic device to implement the method as described in any one of claims 1 to 8.

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

Images