Semiconductor wafer manufacturing AMHS logistics state intelligent prediction method and prediction system

Abstract

The invention provides a semiconductor wafer manufacturing AMHS logistics state intelligent prediction method and system, and the method comprises the steps: carrying out the multi-scale time sequence feature extraction and hierarchical spatial relation modeling through an encoder, and capturing the short-term, middle-term and long-term dynamic characteristics through a multi-scale attention mechanism in the time dimension, a graph attention network GAT and an improved space Transform are adopted in parallel in the spatial dimension to model a local neighborhood and global dependency relationship respectively, and cross-scale spatial-temporal feature fusion is realized through connection of a gating fusion mechanism and a residual error; and decoding the fused complex spatial-temporal features by using a decoder including double-layer convolution decoding and reverse normalization processing, and predicting a logistics state matrix sequence of a plurality of time steps in the future. According to the method, through deep coupling of the multi-scale spatial-temporal features, the problems of spatial-temporal feature separation, incomplete multi-scale information capture and difficult modeling of a complex topological structure in a traditional method are effectively solved, and the accuracy and stability of AMHS logistics state prediction are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing technology, and in particular relates to an intelligent prediction method and system for AMHS logistics status in semiconductor wafer manufacturing. Background Technology

[0002] In the semiconductor manufacturing industry, the Automated Material Handling System (AMHS) is a core infrastructure supporting the efficient operation of wafer manufacturing. It is responsible for the precise and timely transfer of wafer batches between different process equipment, directly impacting the overall capacity, yield, and manufacturing cost of the production line. As semiconductor process nodes continue to shrink, the timeliness and accuracy requirements for material handling in wafer manufacturing are increasing. Accurate prediction of AMHS logistics status has become a key prerequisite for achieving intelligent scheduling and optimization of the production process. By predicting changes in AMHS logistics status in advance, problems such as material congestion and equipment idleness can be effectively avoided, resource allocation can be optimized, and overall production efficiency can be improved.

[0003] However, AMHS logistics status prediction faces complex technical challenges. On the one hand, AMHS logistics status data exhibits significant spatiotemporal coupling characteristics, meaning that changes in the logistics status of a node are not only related to its own historical data but also influenced by changes in the logistics status of surrounding nodes, and this influence relationship evolves dynamically over time. On the other hand, logistics status data contains multi-scale temporal characteristics, exhibiting both short-term fluctuations and periodic change patterns. Furthermore, the physical layout of AMHS forms a complex network topology, where the actual reachability and transmission costs between nodes often depend on the physical connections and layout of transport rails, rather than simple geometric distances. This makes it difficult for traditional methods based on regular grids or Euclidean distance assumptions to effectively model its spatial relationships, further increasing the difficulty of prediction. These characteristics make it difficult for conventional time series forecasting methods to achieve ideal prediction accuracy.

[0004] Existing AMHS (Advanced Semiconductor Manufacturing Services) logistics status prediction methods mainly include traditional statistical methods, machine learning methods, and deep learning techniques. Traditional statistical methods struggle to capture complex nonlinear relationships and spatiotemporal coupling characteristics; machine learning methods, such as support vector machines, while capable of handling a certain degree of nonlinearity, suffer from limitations in computational efficiency and generalization ability when dealing with large-scale, high-dimensional data; existing deep learning techniques, such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), have made some progress in time series prediction, but still have shortcomings in handling multi-scale spatiotemporal features and complex network topologies, making it difficult to fully extract the complex spatiotemporal correlation information in AMHS logistics status data, resulting in prediction accuracy that fails to meet the high-precision requirements of semiconductor manufacturing. Therefore, developing an intelligent prediction method and system for AMHS logistics status in semiconductor wafer manufacturing is of great significance for improving the level of intelligence in semiconductor manufacturing. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent prediction method and system for AMHS logistics status in semiconductor wafer manufacturing, addressing problems in existing technologies such as spatiotemporal feature decoupling, insufficient multi-scale feature capture, and difficulty in spatial relationship modeling, thereby achieving accurate prediction of AMHS logistics status. The adopted technical solution is as follows:

[0006] A method for intelligent prediction of AMHS logistics status in semiconductor wafer manufacturing includes the following steps:

[0007] Step S100: Generate high-dimensional spatiotemporal feature representation:

[0008] Historical logistics data and node network topology data of the AMHS system in semiconductor manufacturing plants are collected and preprocessed. The preprocessed data are then fused with spatiotemporal features to generate a high-dimensional spatiotemporal feature representation.

[0009] Step S200: Perform multi-scale spatiotemporal feature fusion on the high-dimensional spatiotemporal feature representation to obtain fused spatiotemporal features that fuse spatial and temporal features;

[0010] Step S300: Decode and predict the fused spatiotemporal features to obtain the logistics status sequence for multiple future time steps.

[0011] Preferably, the multi-scale spatiotemporal feature fusion in step S200 specifically includes the following steps:

[0012] In the time dimension, short-term, medium-term, and long-term time features are extracted separately through a multi-scale time attention mechanism, and the short-term, medium-term, and long-term time features are fused into multi-scale time features;

[0013] In the spatial dimension, graph attention network and spatial Transformer are used in parallel. Graph attention network combined with local multi-head attention mechanism to obtain local spatial features, and spatial Transformer combined with global multi-head attention mechanism to obtain global spatial features. Then, local spatial features and global spatial features are fused to form a unified spatial representation.

[0014] Multi-scale temporal features and unified spatial representations are fused to obtain the fused spatiotemporal features.

[0015] Preferably, the steps for obtaining multi-scale temporal features specifically include:

[0016] S201. Embed the time features into the corresponding nodes;

[0017] S202. Divide time series data into three time scales: short-term, medium-term, and long-term.

[0018] S203. Apply the attention mechanism to each time scale and calculate the self-attention weights at different time scales.

[0019] S204. Fuse output features from different time scales.

[0020] Preferably, the steps for obtaining a unified spatial representation specifically include:

[0021] S200A, based on high-dimensional spatiotemporal features, extracts local and global spatial features:

[0022] S200B integrates local and global spatial features to form a unified spatial representation.

[0023] Preferably, spatial location embedding is introduced into the global spatial features extracted in S200A.

[0024] Preferably, fusion is performed in S200B through a gating fusion mechanism.

[0025] A semiconductor wafer manufacturing AMHS logistics status intelligent prediction system is used to implement the semiconductor wafer manufacturing AMHS logistics status intelligent prediction method as described above, including:

[0026] The system includes a GUI module, a data preprocessing module, a spatiotemporal coupling module, and a prediction module.

[0027] The GUI module is used to receive historical logistics data and node network topology data from the AMHS system of the semiconductor manufacturing plant in real time, and use them as raw data.

[0028] The data preprocessing module is used to clean, normalize, and construct time features from the collected data;

[0029] The spatiotemporal coupling module is used to capture the temporal dynamic features and spatial relationship features of AMHS logistics status data through multi-scale temporal feature extraction and hierarchical spatial relationship modeling, and to perform spatiotemporal coupling feature fusion.

[0030] The prediction module is used to predict the sequence of logistics state matrices for multiple future time steps through double-layer convolutional decoding and inverse normalization processing, and transmit the generated prediction data to the GUI module.

[0031] Compared with the prior art, the advantages of the present invention are:

[0032] By fusing multi-scale spatiotemporal features, the problem of spatiotemporal feature decoupling in traditional methods is effectively solved, and the dynamic characteristics and complex spatial relationships at different time scales in the AMHS system can be captured simultaneously.

[0033] By using graph attention networks to propagate neighborhood information, the ability to represent spatial relationships is enhanced, enabling better modeling of the topology of AMHS systems.

[0034] By decoding the fused spatiotemporal features through dual-layer convolutional decoding and inverse normalization processing, the accuracy and stability of prediction are improved, enabling effective prediction of logistics status at multiple future time steps.

[0035] The method of this invention can provide accurate logistics status prediction for the AMHS system of semiconductor manufacturing plants, which helps to optimize production scheduling and improve production efficiency. Attached Figure Description

[0036] Figure 1 The overall flowchart of the intelligent prediction method for AMHS logistics status in semiconductor wafer manufacturing provided in one embodiment of the present invention is shown below.

[0037] Figure 2 This is a schematic diagram of the overall architecture of the AMHS intelligent prediction system for semiconductor wafer manufacturing logistics status provided in one embodiment of the present invention.

[0038] Figure 3 This is a schematic diagram of AMHS layout for semiconductor wafer manufacturing provided in one embodiment of the present invention.

[0039] Figure 4 This is a schematic diagram of the time feature extraction module architecture in the intelligent prediction system for AMHS logistics status in semiconductor wafer manufacturing, provided as an embodiment of the present invention.

[0040] Figure 5 This is a schematic diagram of the spatial feature extraction module architecture in an AMHS (Advanced Semiconductor Wafer Manufacturing) logistics status intelligent prediction system provided in an embodiment of the present invention.

[0041] Figure 6This diagram illustrates the comparison between actual and predicted values ​​of the AMHS logistics status intelligent prediction system for semiconductor wafer manufacturing provided in an embodiment of the present invention on the D1 dataset.

[0042] Figure 7 This is an architecture diagram of an AMHS (Advanced Management System for Semiconductor Wafer Manufacturing) logistics status intelligent prediction system provided in one embodiment of the present invention. Detailed Implementation

[0043] The intelligent prediction method and system for AMHS logistics status in semiconductor wafer manufacturing of the present invention will be described in more detail below with reference to the schematic diagrams, which illustrate preferred embodiments of the present invention. It should be understood that those skilled in the art can modify the present invention described herein while still achieving the advantageous effects of the present invention. Therefore, the following description should be understood as being of general knowledge to those skilled in the art and is not intended to limit the present invention.

[0044] like Figures 1-6 A method for intelligent prediction of AMHS logistics status in semiconductor wafer manufacturing includes the following steps:

[0045] Step S100: Generate a high-dimensional spatiotemporal feature representation.

[0046] Historical logistics data and node network topology data of the AMHS system in semiconductor manufacturing plants are collected and preprocessed. The preprocessed data are then fused with spatiotemporal features to generate a high-dimensional spatiotemporal feature representation with dimensions [B,N,T,d].

[0047] AMHS mainly consists of transport tracks and transport carts (OHT). The transport tracks are used to connect different wafer processing equipment, and all of them are unidirectional tracks, such as... Figure 3 As shown, the OHT travels on transport rails, enabling direct handling of wafer cards at the exit of wafer processing equipment, and is the transport tool in AMHS.

[0048] Historical logistics status data refers to the logistics flow of each node (i.e., the starting point of the transport guide) in AMHS within a specific time step. Logistics flow refers to the number of OHTs passing through a specific node within a time step.

[0049] The topological relationship data between nodes is represented by an adjacency matrix. Description: Matrix element A[i,j] represents the rail connection relationship and physical distance from node i to node j. Since the rail system operates in one direction, A is an asymmetric matrix. When there is a rail connection between nodes, the value is the physical distance between the two nodes; when there is no connection, the value is 0.

[0050] High-dimensional spatiotemporal features refer to the feature representations formed by integrating temporal features and spatial topological information from the original single-dimensional historical logistics data and raising it to a preset feature dimension d, which can carry richer spatiotemporal pattern semantics. They are usually represented by tensors of shape [B,N,T,d], where B is the batch size (the number of samples processed simultaneously during one forward and backward propagation of the model during training), N is the number of nodes, T is the time step, and d is the feature dimension.

[0051] Preprocessing includes data cleaning, time alignment, and normalization, and the processed data is used to build training, validation, and testing datasets.

[0052] The steps for obtaining high-dimensional spatiotemporal feature representations are existing technologies, and include steps S101 to S103.

[0053] Step S101: Extract discrete time attributes such as minutes, hours, days of the week, dates, and months based on the timestamps of the input sequence (the original input data), and form a time feature vector through embedding mapping; align the time feature with the input sequence according to the time dimension, and extend it to all nodes to obtain the time feature tensor;

[0054] Step S102: Read the topological relationship data between the guide rails of the nodes and generate an adjacency matrix. Where A[i,j] represents the guide rail connection relationship and physical distance from node i to node j; then, a Gaussian kernel function is used:

[0055]

[0056] Normalize the non-zero distances, where σ is the standard deviation of all non-zero distances, to obtain the weighted adjacency matrix. As a constraint input characterizing the spatial topology.

[0057] Step S103: Using the normalized historical logistics data as the basic input, the data is fused through two stages: time feature injection and spatial feature injection.

[0058] First, the temporal feature tensor generated in step S101 is added to the basic input in the feature dimension through the temporal location encoding module to achieve deep fusion of temporal information; at the same time, the weighted adjacency matrix obtained in step S102 is linearly transformed into a spatial feature vector through the spatial location embedding module, and then added to the basic input in the feature dimension to achieve fusion of spatial topological information.

[0059] The above injection operation ensures that the feature dimension of the output of each module remains consistent at d, and finally forms a unified high-dimensional spatiotemporal feature representation with dimension [B,N,T,d], which serves as the input of the multi-scale spatiotemporal feature fusion module in step S200.

[0060] Step S200: Perform multi-scale spatiotemporal feature fusion on the high-dimensional spatiotemporal feature representation to obtain fused spatiotemporal features that fuse spatial and temporal features.

[0061] By fusing multi-scale spatiotemporal features, the problem of spatiotemporal feature decoupling in traditional methods is effectively solved, and the dynamic characteristics and complex spatial relationships at different time scales in the AMHS system can be captured simultaneously.

[0062] By using graph attention networks to propagate neighborhood information, the ability to represent spatial relationships is enhanced, enabling better modeling of the topology of AMHS systems.

[0063] Step S200 specifically includes the following three parts:

[0064] First, short-term, medium-term, and long-term time features are extracted separately using a multi-scale time attention mechanism in the time dimension, and then the short-term, medium-term, and long-term time features are fused into multi-scale time features.

[0065] like Figure 4 As shown, the specific steps for obtaining multi-scale temporal features include:

[0066] S201. Embed the time feature into the corresponding node (i.e., the starting point of the corresponding transport rail). Specifically, this includes:

[0067] Based on the timestamp of the input data, discrete time attributes such as minutes, hours, weekdays, dates, and months are extracted and mapped into vector representations to form multi-class time attribute embeddings.

[0068] The input data refers to the original dataset containing logistics flow observations and their corresponding timestamps, which is organized into batches after preprocessing for model training.

[0069] Temporal features refer to discrete attribute values ​​extracted from timestamps that can reflect periodic patterns, and are used to provide temporal context information for the model.

[0070] The embeddings of multiple time attributes are aligned with the input sequence features in the time dimension and then concatenated to obtain the time feature embedding matrix;

[0071] The temporal feature embedding matrix is ​​then extended along the spatial dimension and copied to all nodes, thus forming an extended temporal feature embedding tensor.

[0072] The temporal feature embedding matrix refers to the two-dimensional matrix representation formed by mapping discrete time attributes to continuous vectors through an embedding layer and then fusing them. Its dimension is [T×d], where T is the time step and d is the feature dimension.

[0073] The purpose of step S201 is to establish the correspondence between time features and nodes, that is, to find the time features corresponding to each node.

[0074] S202. Divide time series data into three time scales: short-term, medium-term, and long-term.

[0075] Time series data refers to a spatiotemporal feature tensor formed by fusing logistics flow characteristics and time characteristics. Its shape is [B,N,T,d], where B is the batch size, N is the number of nodes, T is the time step, and d is the feature dimension.

[0076] The short-term time scale refers to the time series segment of the most recent 6 time steps (corresponding to the most recent 3 minutes), the medium-term time scale refers to the time series segment of the most recent 7th to 12th time steps (corresponding to 3 to 6 minutes ago), and the long-term time scale refers to the time series segment of the 1st to (T-12th)th time steps (corresponding to 6 minutes ago).

[0077] S201 provides time-encoded feature representations for S202. Only by first converting discrete-time attributes into continuous vectors (S201) can multi-scale partitioning and attention computation be performed on these vectors (S202-S204). This sequence ensures that the model can utilize temporal context information to analyze changes in logistics status at different time scales.

[0078] Specifically, multi-scale time series modeling operations are performed, and short-term, medium-term, and long-term time dependencies are captured through a multi-scale time attention mechanism.

[0079] S203. Apply the attention mechanism to each time scale separately to calculate the self-attention weights at different time scales. Specifically, this includes:

[0080] A1: Project the time series input at each time scale into a query matrix Q, a key matrix K, and a value matrix V; the time series input refers to the feature segments of short-term, medium-term, or long-term time scales extracted from S202;

[0081] A2: Calculate the self-attention head at the short-term, medium-term, and long-term scales respectively:

[0082]

[0083] A3: Attention weight calculation:

[0084]

[0085] S204. Fuse output features from different time scales.

[0086] Output features refer to the feature representations at short-term, medium-term, and long-term time scales after processing by S203. In S203, attention weights (output of the softmax function) are used to perform a weighted summation on the value matrix V, thereby generating new feature representations.

[0087] Therefore, the weights calculated by S203 directly determine how information is extracted and combined from the original features to generate more representative output features. Specifically, for each time step, the attention weights determine which other time steps' feature information within the same scale should be focused on at that time step, thereby generating new features that incorporate contextual relevance.

[0088] Feature fusion involves filling / stitching and feedforward fusion of the outputs from the three time scales to obtain unified multi-scale time features.

[0089] Second, graph attention networks and spatial Transformers are used in parallel in the spatial dimension. The graph attention network is combined with a local multi-head attention mechanism to obtain local spatial features, and the spatial Transformer is combined with a global multi-head attention mechanism to obtain global spatial features. Then, the local spatial features and global spatial features are fused to form a unified spatial representation.

[0090] like Figure 5 As shown, the specific steps for obtaining a unified spatial representation include:

[0091] S200A, based on high-dimensional spatiotemporal features, performs hierarchical spatial relationship modeling operations to extract local and global spatial features.

[0092] High-dimensional spatiotemporal features refer to the feature tensors with dimensions [B,N,T,d] generated after processing in step S100, which contain fused time information and logistics flow information.

[0093] Hierarchical spatial relationship modeling refers to using different modeling strategies to handle local neighborhood relationships (based on physical connections) and global node relationships (based on feature similarity) respectively, in order to fully capture the complex spatial dependency structure of the AMHS network.

[0094] Spatial location embedding is introduced during the extraction of global spatial features.

[0095] (1) Local spatial feature extraction:

[0096] We employ a Graph Attention Network (GAT) and a local multi-head attention mechanism in parallel to extract neighborhood structure dependencies based on physical connections of guide rails and local contextual patterns based on feature similarity, respectively.

[0097] Node features refer to the d-dimensional feature vector corresponding to each node at a specific time step, representing the logistics status code of that node at that moment.

[0098] Local contextual patterns refer to the correlation and interaction patterns between node features within a local neighborhood (between directly connected nodes).

[0099] Linear transformation of input node features is performed based on graph attention network;

[0100] Calculate the attention coefficients for nodes i and j:

[0101] α ij =softmax j (LeakyReLU(a T [Wh i ||Wh j ]))

[0102] Where, α ij Here, a and W are the attention coefficients, and h is the trainable parameter. i and h j These are node features. The domain features are then weighted and aggregated based on this attention weight:

[0103]

[0104] Where σ is the activation function, and N(i) represents the neighborhood set of node i.

[0105] Simultaneously, based on the local multi-head attention mechanism, a query, key, and value matrix is ​​constructed within the local context, attention is calculated, and the outputs of each head are concatenated to obtain another set of local context representations h. l-mha .

[0106] Output h′ of the graph attention network i With local multi-head attention output h l-mha By fusing the data, we obtain the local spatial features h. local .

[0107] (2) Global spatial feature extraction:

[0108] We employ a spatial Transformer and a global multi-head attention mechanism in parallel to model long-distance dependencies across nodes.

[0109] Spatial location embedding is a key component for realizing spatially-aware modeling, and it is constructed by using a weighted adjacency matrix that represents the network topology. (Where N is the number of nodes, and matrix element W[i,j] represents the spatial relationship strength from node i to node j) is taken as input and mapped to a spatial location embedding matrix through a learnable linear transformation layer Linear(N,d). Where d is the preset feature dimension. During training, this linear transformation layer learns to encode the topological structure as a continuous vector representation.

[0110] Specifically, for each node i, its spatial location embedding vector It is obtained from the i-th row of the adjacency matrix (representing the connection relationship between node i and all other nodes in the network) through a linear transformation, that is:

[0111] e i =Linear(W[i,:])

[0112] In practical implementation, the calculated spatial location embedding matrix E is used. spatial Expanding along the batch dimension (B copies) and the time dimension (T copies) forms a spatial location embedding tensor with the same shape as the input features. Subsequently, the embedded tensor is fused with the original spatiotemporal input features by element-wise addition:

[0113] X enhanced =X+E entended

[0114] This allows each node to be injected with its unique spatial topology information at each time step, providing structural bias for the subsequent attention mechanism.

[0115] Based on an improved spatial Transformer, spatial location embeddings are first added to the input features to enhance the model's perception of network topology. Then, the enhanced features are projected into query, key, and value representations.

[0116] Q = X enhanced W Q K = X enhanced W K V = X enhanced W V

[0117] Calculate global dependencies using an improved spatial Transformer:

[0118]

[0119] This formula captures the global dependencies between all nodes, yielding the output h of the spatial Transformer. str .

[0120] Based on a global multi-head attention mechanism, a query, key, and value matrix is ​​constructed within the global context. Attention is calculated, and the outputs of each head are concatenated to obtain another set of global context representations h. g-mha .

[0121] The global context refers to the overall environment comprised of all nodes in the entire AMHS network, reflecting the global distribution of logistics status and interaction patterns.

[0122] Global context representation refers to a feature representation that incorporates information from all nodes and is computed through a global attention mechanism.

[0123] Output h from the spatial Transformer str The output h of the global multi-head attention mechanism g-mha By fusing the data, we obtain the global spatial features h. global This allows for the simultaneous capture of long-range dependencies across nodes and global patterns.

[0124] S200B integrates local and global spatial features to form a unified spatial representation.

[0125] Specifically, a gating fusion mechanism is used to analyze local spatial features h. local With global spatial features h global Adaptive weighted fusion is performed to obtain a unified spatial representation h. space .

[0126] III. Integrating multi-scale temporal features with a unified spatial representation h space Fusion to obtain fusion spatiotemporal characteristics.

[0127] Step S300: Decode and predict the fused spatiotemporal features to obtain a sequence of logistics status for multiple future time steps. This specifically includes steps S301 to S302.

[0128] The fused spatiotemporal features output from step S200 are input into the prediction layer for decoding and mapping.

[0129] S301: High-dimensional spatiotemporal features are projected onto the prediction output space through two-layer convolutional decoding.

[0130] The prediction layer contains two cascaded convolution operations:

[0131] Temporal Dimension Mapping Convolution: Using a 1×1 convolution kernel, dimensionality reduction mapping is performed in the temporal dimension, compressing the input time step T to the prediction time step T. P :

[0132] Conv1:

[0133] Feature Dimension Reduction Convolution: Using a 1×1 convolution kernel, dimensionality reduction is performed along the feature dimension, reducing the feature dimension d to the number of output channels C (C=1, corresponding to a single value of logistics flow).

[0134] Conv2:

[0135] After two layers of convolutional decoding, the predicted representation tensor for multiple future time steps is obtained:

[0136]

[0137] Among them, T P For the prediction time steps (T) P =6, corresponding to the next 3 minutes), N is the number of nodes, and B is the output feature dimension.

[0138] S302: Perform inverse normalization on the prediction results to obtain the actual logistics flow value.

[0139] The prediction tensor obtained from convolution decoding After performing inverse normalization, using the mean μ and standard deviation σ saved during the data preprocessing stage, the predicted results are restored to the actual logistics flow using the following formula:

[0140]

[0141] The inversely normalized prediction matrix is ​​expressed as follows:

[0142]

[0143] This matrix directly reflects the future T of semiconductor AMHS systems. P The logistics status of each node at each time step (corresponding to the next 3 minutes), that is, the number of OHTs passing through each node per unit time.

[0144] In summary, by using convolution and inverse normalization, the accuracy and stability of predictions are improved, enabling effective prediction of logistics status at multiple future time steps. This provides accurate logistics status predictions for the AMHS system in semiconductor manufacturing plants, helping to optimize production scheduling and improve production efficiency.

[0145] To verify the beneficial effects of this invention, scientific demonstration was conducted through experiments.

[0146] The effectiveness of the present invention, the Semiconductor Wafer Manufacturing AMHS Logistics Status Intelligent Prediction System (MSTC-Model) based on multi-scale spatiotemporal coupling, was verified using a logistics dataset from an automated material handling system in a semiconductor smart factory in Shanghai. The dataset contains real-time logistics data for each OHT transport path in the semiconductor smart factory, covering the flow information of the transport rails within one month.

[0147] Table 1 summarizes the information of the dataset; the data is collected every 30 seconds, covering different production load scenarios, including peak and off-peak periods, and all data has been standardized (normalized to the [0,1] interval).

[0148] Table 1. Dataset Description

[0149]

[0150] In Table 1, OHT refers to the transport trolley. The OHT travels on the transport rail and is the transport vehicle in AMHS. The node is the starting point of the transport rail.

[0151] The Transformer encoder and decoder have 1 layer, convolutional layers 2, attention heads 4, model dimension 64, learning rate 0.001, and Dropout probability 0.2. The optimizer is Adam, with 100 training epochs and a batch size of 8. Early stopping is used to prevent overfitting; training stops when the validation set MAE does not decrease for 10 consecutive epochs.

[0152] To verify the model's effectiveness, mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are used. The specific formulas are as follows:

[0153]

[0154] Where n is the number of samples, y i For the i-th actual observation, Let be the i-th predicted value.

[0155] To verify the effectiveness of the invented MSTC-Model model, experiments were conducted using the D1 dataset under the same experimental configuration, and it was compared with VAR (Vector Autoregression), SVR (Support Vector Regression), DCRNN (Diffusion Convolutional Recurrent Neural Network), and GWNET (Graph Neural Network with Diffusion Convolution and Self-Attention Mechanism).

[0156] The experimental results were compared and analyzed using the three performance evaluation indicators mentioned above. The results are shown in Table 2. The MSTC-Model method provided by this invention is superior to other methods in all three performance evaluation indicators.

[0157] Table 2 Performance comparison of the present invention with other methods

[0158]

[0159] To more intuitively demonstrate the effectiveness of the MSTC-Model method, a key node (i.e., the starting point of the critical transport rail) is selected from the D1 dataset to visualize the prediction results, and the predicted values ​​are compared with the actual values, such as... Figure 6 As shown, the model provided by this invention can make good predictions and effectively fit the changing trend of material flow, further verifying the effectiveness of the model.

[0160] Reference Figure 7 A semiconductor wafer manufacturing plant logistics congestion prediction system includes a GUI module, a data preprocessing module, a spatiotemporal coupling module, and a prediction module.

[0161] The GUI module provides a user interface, receives real-time logistics status data collected from semiconductor manufacturing plants, and supports the visualization of prediction results and congestion alarm information.

[0162] The data preprocessing module is used to process the collected historical logistics data. Its functions include data cleaning, time alignment, data normalization, and construction of time features. Simultaneously, it generates an adjacency matrix based on the topological relationships of the factory's transport rails and calculates edge weights using a Gaussian kernel function to obtain a weighted adjacency matrix for spatial modeling.

[0163] The spatiotemporal coupling module extracts temporal features through a multi-scale temporal attention mechanism and models local and global spatial features in parallel using a graph attention network and a spatial Transformer. After gating and fusion, the temporal and spatial features are fused, and residual connections and layer normalization are combined to obtain a cross-scale spatiotemporal representation.

[0164] The prediction module is used to predict the sequence of logistics state matrices for multiple future time steps through double-layer convolutional decoding and inverse normalization processing, and transmits the generated prediction data to the GUI module.

[0165] It should be noted that the functional modules of this system can be implemented either through hardware circuits or by loading software programs onto a computer or server. The GUI module, data preprocessing module, spatiotemporal coupling module, and prediction module can operate as independent functional units or be integrated and deployed on the same platform. This system can realize intelligent logistics prediction and congestion alarm for semiconductor AMHS systems, improving the stability and efficiency of factory production operations.

[0166] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0167] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0168] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0169] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0170] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A method for intelligent prediction of AMHS logistics status in semiconductor wafer manufacturing, characterized in that, Includes the following steps: Step S100: Generate high-dimensional spatiotemporal feature representation: Historical logistics data and node network topology data of the AMHS system in semiconductor manufacturing plants are collected and preprocessed. The preprocessed data are then fused with spatiotemporal features to generate a high-dimensional spatiotemporal feature representation. Step S200: Perform multi-scale spatiotemporal feature fusion on the high-dimensional spatiotemporal feature representation to obtain fused spatiotemporal features that fuse spatial and temporal features; Step S300: Decode and predict the fused spatiotemporal features to obtain the logistics status sequence for multiple future time steps.

2. The intelligent prediction method for AMHS logistics status in semiconductor wafer manufacturing according to claim 1, characterized in that, The multi-scale spatiotemporal feature fusion in step S200 specifically includes the following steps: In the time dimension, short-term, medium-term, and long-term time features are extracted separately through a multi-scale time attention mechanism, and the short-term, medium-term, and long-term time features are fused into multi-scale time features; In the spatial dimension, graph attention network and spatial Transformer are used in parallel. Graph attention network combined with local multi-head attention mechanism to obtain local spatial features, and spatial Transformer combined with global multi-head attention mechanism to obtain global spatial features. Then, local spatial features and global spatial features are fused to form a unified spatial representation. Multi-scale temporal features and unified spatial representations are fused to obtain the fused spatiotemporal features.

3. The intelligent prediction method for AMHS logistics status in semiconductor wafer manufacturing according to claim 2, characterized in that, The specific steps for obtaining multi-scale temporal features include: S201. Embed the time features into the corresponding nodes; S202. Divide time series data into three time scales: short-term, medium-term, and long-term. S203. Apply the attention mechanism to each time scale and calculate the self-attention weights at different time scales. S204. Fuse output features from different time scales.

4. The intelligent prediction method for AMHS logistics status in semiconductor wafer manufacturing according to claim 2, characterized in that, The specific steps for obtaining a unified spatial representation include: S200A, based on high-dimensional spatiotemporal features, extracts local and global spatial features: S200B integrates local and global spatial features to form a unified spatial representation.

5. The intelligent prediction method for AMHS logistics status in semiconductor wafer manufacturing according to claim 4, characterized in that, S200A introduces spatial location embedding into the extraction of global spatial features.

6. The intelligent prediction method for AMHS logistics status in semiconductor wafer manufacturing according to claim 4, characterized in that, In S200B, fusion is achieved through a gating fusion mechanism.

7. A semiconductor wafer manufacturing AMHS logistics status intelligent prediction system, used to implement the semiconductor wafer manufacturing AMHS logistics status intelligent prediction method as described in any one of claims 1 to 6, characterized in that, include: The system includes a GUI module, a data preprocessing module, a spatiotemporal coupling module, and a prediction module. The GUI module is used to receive historical logistics data and node network topology data from the AMHS system of the semiconductor manufacturing plant in real time, and use them as raw data. The data preprocessing module is used to clean, normalize, and construct time features from the collected data; The spatiotemporal coupling module is used to capture the temporal dynamic features and spatial relationship features of AMHS logistics status data through multi-scale temporal feature extraction and hierarchical spatial relationship modeling, and to perform spatiotemporal coupling feature fusion. The prediction module is used to predict the sequence of logistics state matrices for multiple future time steps through double-layer convolutional decoding and inverse normalization processing, and transmit the generated prediction data to the GUI module.

Images