Semiconductor wafer manufacturing factory logistics jam prediction method and system

Abstract

The invention discloses a semiconductor wafer manufacturing factory logistics jam prediction method and system, and relates to the technical field of semiconductor manufacturing process optimization, and the method comprises the steps: collecting and preprocessing historical logistics data of a semiconductor manufacturing intelligent factory, carrying out the spatial-temporal feature fusion of the preprocessed data, generating high-dimensional representation data, and carrying out the prediction of the logistics jam of the semiconductor wafer manufacturing factory. Performing coding feature extraction processing on the high-dimensional representation data, extracting spatio-temporal features through a time sequence dynamic sensing operation and a spatial relationship self-adaptive operation, performing spatio-temporal feature fusion processing, performing cross-sequence feature integration on a processing result and complex spatio-temporal features, and performing decoding prediction processing on the integrated features, and through a causal constraint attention mechanism and cross-sequence correlation modeling, predicting a logistics state matrix sequence, and based on the logistics state matrix sequence, carrying out congestion state determination. According to the method, high-precision prediction and quick response of logistics blockage of the semiconductor factory are realized, and the problems of spatial-temporal characteristic splitting, response lag and high false alarm rate of a traditional method are solved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing process optimization technology, specifically to a method and system for predicting logistics congestion in semiconductor wafer manufacturing plants. Background Technology

[0002] With the widespread application of smart factories for 12-inch semiconductor manufacturing, automated material handling systems (AWTS) have become a key component of semiconductor manufacturing. AWTS uses overhead cranes to transport wafers on complex track systems, ensuring efficient transfer of wafers between different processing areas. However, due to the complex guide rail layout, large-scale overhead cranes, dynamic random handling tasks, and regionally differentiated operating speeds, AWTS frequently experience temporary congestion during operation. This congestion not only reduces the overall handling efficiency of the AWTS but may also lead to longer waiting times for production equipment, decreased capacity utilization, and even affect product yield. Therefore, predicting the logistics of each area of ​​the AWTS in advance to identify potential critical congestion nodes and time periods is of great significance for optimizing scheduling strategies and improving transportation efficiency.

[0003] In the field of logistics congestion prediction in smart semiconductor manufacturing factories, traditional methods mainly include statistical and machine learning-based approaches. Statistical methods, such as spatiotemporal moving average autoregressive models and vector autoregressive models, can consider spatiotemporal correlations, but they struggle to accurately describe the complex nonlinear relationships between multiple factors in logistics, are highly dependent on the accuracy of model parameters, and are greatly affected by noise statistics. Machine learning methods, such as improved K-nearest neighbor models and adaptive multi-kernel support vector machines, while overcoming some shortcomings of statistical methods, struggle to effectively describe the complex relationships and dynamic changes between features when facing large-scale complex traffic networks, resulting in weak model adaptability and generalization ability. In recent years, deep learning methods have been increasingly applied to logistics prediction, such as diffusing convolutional recurrent neural networks and long short-term memory networks. These methods can better consider the complex relationships and dynamic changes between large-scale features, exhibiting good adaptability and generalization ability. However, most existing deep learning methods fail to fully consider the dependencies between long-distance time-series data in traffic networks, and their scalability is insufficient, affecting the accuracy and generalization ability of the prediction models. Summary of the Invention

[0004] In view of the above-mentioned existing problems, the present invention provides a method and system for predicting logistics congestion in semiconductor wafer manufacturing plants, in order to solve the problems of temporary OHT congestion, extended waiting time for production equipment, decreased capacity utilization, insufficient scalability, and reduced accuracy and generalization ability of prediction models in the prior art.

[0005] To address the aforementioned technical problems, a method for predicting logistics congestion in semiconductor wafer manufacturing plants is proposed, including:

[0006] Historical logistics data from a smart semiconductor manufacturing factory is collected and preprocessed. The preprocessed data undergoes spatiotemporal feature fusion to generate high-dimensional representation data. Encoded feature extraction is performed on the high-dimensional representation data. Spatiotemporal features are extracted through time-series dynamic perception and spatial relationship adaptive operations, and spatiotemporal feature fusion is performed. The processing results are integrated with the complex spatiotemporal features output by encoding through cross-sequence feature integration. The integrated features are then decoded and predicted. The logistics state matrix sequence is predicted through causal constraint attention mechanism and cross-sequence association modeling, and congestion status is determined based on the logistics state matrix sequence.

[0007] As a preferred embodiment of the semiconductor wafer manufacturing plant logistics congestion prediction method of the present invention, the step of performing spatiotemporal feature fusion includes generating periodic position codes for the time dimension and extending them to the spatial dimension.

[0008] Generate topological structure encoding for the spatial dimension, enhance spatial relationship representation through neighborhood information propagation, and extend the results to the temporal dimension;

[0009] The encoding representations of the time and space dimensions are combined with multidimensional features.

[0010] As a preferred embodiment of the semiconductor wafer manufacturing plant logistics congestion prediction method of the present invention, the encoded feature extraction process includes performing a time-series dynamic perception operation and capturing time dynamic characteristics through local context feature extraction.

[0011] Perform spatial relationship adaptive operation, dynamically generate relationship strength matrix based on node features and adjust spatial connection weights;

[0012] The processing results are then subjected to feature stabilization.

[0013] Enhance feature representation capabilities through nonlinear feature enhancement.

[0014] As a preferred embodiment of the semiconductor wafer manufacturing plant logistics congestion prediction method of the present invention, the decoding prediction process includes: utilizing a causal constraint attention mechanism and employing unidirectional information flow control to ensure temporal causality;

[0015] Implement cross-sequence association modeling to establish the association between the decoded sequence and the encoded output;

[0016] Adaptive operations combining spatial relationships are used to process spatial features;

[0017] Output the predicted sequence with multiple nodes and multiple time steps.

[0018] As a preferred embodiment of the semiconductor wafer manufacturing plant logistics congestion prediction method of the present invention, the generation of periodic position codes includes: calculating trigonometric function position vectors based on time step indices, constructing a time-dimensional coding matrix, and extending the coding matrix along the spatial dimension;

[0019] Constructing the time-dimensional encoding matrix includes the shape of the input data. For each time step index, the vector is calculated to generate a time dimension encoding matrix according to the formula, which is expressed as:

[0020]

[0021] Among them, E TP The embedding matrix is ​​a time position matrix, where t is the time step index and d is the dimension index. k The total dimension of the vector. Where T is the frequency adjustment factor, T' is the number of time steps, N is the number of nodes, and C is the feature dimension;

[0022] Extending the encoding matrix along the spatial dimension includes E TP The copy is extended and replicated across all nodes to form an extended temporal location embedding tensor.

[0023] The generated topology encoding includes generating an initial embedding matrix for each node, performing Laplacian smoothing through multi-layer GCN operations, aggregating neighboring node information, and each layer including adjacency matrix normalization and feature transformation operations.

[0024] The formula for generating the initial embedding matrix is ​​expressed as:

[0025]

[0026] The formula for operating a multi-layer GCN is expressed as follows:

[0027]

[0028] in, For the initial spatial embedding of node i, W s For trainable weights, b s For trainable bias, s i Let be the static feature vector of node i. Let A be the spatial embedding matrix of the l-th layer, and let I be the adjacency matrix. N It is the identity matrix. To add the adjacency matrix after adding self-loops, W is the degree matrix. (l) Here, σ represents the weights of the l-th layer GCN, and σ is the activation function. Let be the spatial embedding matrix of the (l-1)th layer.

[0029] As a preferred embodiment of the semiconductor wafer manufacturing plant logistics congestion prediction method of the present invention, the local context feature extraction includes: projecting the input features into query, key, and value representations; performing sliding window feature extraction on the query and key; calculating attention weights and weighted fusion of value features; and performing multiple sets of processing in parallel and merging the results.

[0030] The relationship strength matrix includes calculating the similarity measure between node features, generating dynamic correlation weights through a normalized exponential function, and adjusting the static connection matrix and the dynamic weight matrix element by level.

[0031] The cross-sequence feature integration includes applying unidirectional information flow control to the decoder query sequence, applying global context extraction to the encoder key sequence, calculating cross-sequence attention weights, and fusing features.

[0032] As a preferred embodiment of the semiconductor wafer manufacturing plant logistics congestion prediction method of the present invention, the congestion status determination includes: extracting the speed index of each node from the prediction sequence, and comparing the speed index with dynamic threshold conditions to determine the congestion status.

[0033] The beneficial effects of this preferred technical solution are that by extracting the speed index of each node from the prediction sequence, the complex logistics status prediction results are transformed into a quantifiable and intuitive indicator. Furthermore, by comparing the speed index with dynamic threshold conditions to determine the congestion status, a context-based judgment standard is added, which improves the accuracy and adaptability of the judgment results and avoids false alarms and missed alarms.

[0034] As a preferred embodiment of the semiconductor wafer manufacturing plant logistics congestion prediction system of the present invention, it is characterized by including a GUI module, a spatiotemporal information input module, a Transformer encoder module, and a Transformer decoder module.

[0035] The GUI module is used to receive real-time logistics data from the semiconductor manufacturing smart factory collected by the user, including flow rate, occupancy rate and average speed index information, and use the data as raw data.

[0036] The spatiotemporal information input module is used to convert the collected logistics data into a high-dimensional representation containing spatiotemporal features, and enhance the model's ability to perceive spatiotemporal information through a spatiotemporal location embedding strategy.

[0037] The Transformer encoder module is used to capture the temporal and spatial dynamic features of logistics data through a time-aware multi-head self-attention layer and a dynamic graph convolutional layer.

[0038] The Transformer decoder module is used to predict the future multi-time step logistics state matrix sequence by using an autoregressive approach, combining masked time-aware multi-head self-attention and cross-sequence time-aware multi-head self-attention. It also determines whether a node is congested by using the corresponding speed exponent magnitude. Finally, the generated prediction data is transmitted to the GUI module for path planning.

[0039] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a method for predicting material flow congestion in a semiconductor wafer manufacturing plant.

[0040] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for predicting material flow congestion in a semiconductor wafer manufacturing plant.

[0041] The beneficial effects of this invention are as follows: By optimizing the core calculation process, this invention enables logistics forecasting to be completed in a shorter time and allows the model to automatically adapt to different production load environments, solving the problem that calculation time increases rapidly with the complexity of the problem, and significantly improving the timeliness and environmental adaptability of forecasts; by comprehensively analyzing the temporal variation patterns and spatial correlation patterns in logistics data, it achieves a comprehensive grasp of complex spatiotemporal correlations, overcoming the problem of single modeling in existing technologies and effectively improving forecast accuracy; by incorporating time sequence perception and spatial location difference information into the forecasting process, and considering the unique physical constraints of automated material handling, the model is made more in line with actual scenarios, achieving a more scientific and reasonable effect; by establishing a working mode that combines real-time monitoring and future forecasting, it enables the early identification of potential logistics congestion, providing a basis for optimizing scheduling strategies and improving the operating efficiency and production efficiency of automated material handling systems. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the overall process of a semiconductor wafer manufacturing plant logistics congestion prediction method according to an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the overall architecture of the prediction model for a semiconductor wafer manufacturing plant logistics congestion prediction method provided in one embodiment of the present invention.

[0045] Figure 3This diagram illustrates the evaluation metrics of different models of a semiconductor wafer manufacturing plant logistics congestion prediction method provided in an embodiment of the present invention across 20 time steps on the D3 dataset.

[0046] Figure 4 This diagram illustrates the comparison between the actual and predicted values ​​of the prediction model for a semiconductor wafer manufacturing plant logistics congestion prediction method provided in an embodiment of the present invention on the D3 dataset.

[0047] Figure 5 A system flowchart of a semiconductor wafer manufacturing plant logistics congestion prediction system provided in one embodiment of the present invention. Detailed Implementation

[0048] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0049] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for predicting logistics congestion in a semiconductor wafer manufacturing plant is provided, comprising:

[0050] S100: Collects historical logistics data from smart semiconductor manufacturing factories and preprocesses it. Then, it fuses the spatiotemporal features of the preprocessed data to generate high-dimensional representation data.

[0051] S200: Encodes and extracts features from high-dimensional representation data. It extracts spatiotemporal features through time-series dynamic perception and spatial relationship adaptive operations, and performs spatiotemporal feature fusion processing. The processing results are then integrated with the complex spatiotemporal features of the encoded output across sequences.

[0052] S300: Decodes and predicts the integrated features, predicts the logistics state matrix sequence through causal constraint attention mechanism and cross-sequence association modeling, and determines the congestion status based on the logistics state matrix sequence.

[0053] It should be noted that by processing and merging multiple sets of data, the robustness of data processing and the comprehensiveness of prediction results are improved; the calculation of the relationship strength matrix realizes the dynamic adaptive adjustment of the correlation between nodes, enabling the model to better capture the dynamic changes in logistics; cross-sequence feature integration effectively integrates local and global information, enhancing the model's ability to understand complex logistics scenarios; and the comparison between the speed index and dynamic threshold in the congestion status determination improves the accuracy and timeliness of prediction, avoiding false alarms or missed alarms that may be caused by fixed thresholds. Ultimately, the goal of accurate and forward-looking congestion prediction for complex and dynamic logistics systems is achieved, which has significant technological progress and practical value.

[0054] Example 2, refer to Figure 1 and Figure 2 This is a second embodiment of the present invention, which provides a method for predicting logistics congestion in semiconductor wafer manufacturing plants, including:

[0055] In step S100, the historical logistics data includes flow rate, occupancy rate, and average speed index; the preprocessing includes data cleaning, time alignment, and data standardization, and the processed data is used to construct a dataset.

[0056] In this embodiment of the application, step S100, the spatiotemporal feature fusion includes steps S101 to S103:

[0057] S101: Generate periodic location codes for the time dimension and extend them to the spatial dimension;

[0058] S102: Generate topological structure encoding for the spatial dimension, enhance spatial relationship representation through neighborhood information propagation, and extend the results to the time dimension;

[0059] S103: Combine the encoded representations of the time dimension and the spatial dimension into multidimensional features.

[0060] In step S101, generating the periodic position code includes steps A1 to A3:

[0061] A1: Calculate the trigonometric function position vector based on the time step index;

[0062] A2: Construct the time-dimensional encoding matrix; where constructing the time-dimensional encoding matrix includes the shape of the input data. For each time step index, the vector is calculated to generate a time dimension encoding matrix according to the formula, which is expressed as:

[0063]

[0064] Among them, E TP The embedding matrix is ​​a time position matrix, where t is the time step index and d is the dimension index.k The total dimension of the vector. Here, T' is the frequency adjustment factor, N is the number of time steps, N is the number of nodes, and C is the feature dimension.

[0065] A3: and extend the encoding matrix along the spatial dimension; extending the encoding matrix along the spatial dimension includes, E TP The copy is extended and replicated across all nodes to form an extended temporal location embedding tensor.

[0066] In step S102, the generation of topology structure encoding includes steps B1 to B2:

[0067] B1: Generate an initial embedding matrix for each node and perform Laplacian smoothing through multiple GCN layers; the formula for generating the initial embedding matrix is ​​expressed as:

[0068]

[0069] in, For the initial spatial embedding of node i, W s For trainable weights, b s For trainable bias, s i Let i be the static feature vector of node i;

[0070] The formula for operating a multi-layer GCN is expressed as follows:

[0071]

[0072] in, Let A be the spatial embedding matrix of the l-th layer, and let I be the adjacency matrix. N It is the identity matrix. To add the adjacency matrix after adding self-loops, W is the degree matrix. (l) Here, σ represents the weights of the l-th layer GCN, and σ is the activation function. Let be the spatial embedding matrix of the (l-1)th layer.

[0073] B2: Aggregates neighboring node information, each layer includes adjacency matrix normalization and feature transformation operations; after propagation through L layers, it outputs the final spatial embedding matrix; the spatial embedding matrix is ​​then... The copy is extended across all time steps to form an extended spatial location embedding tensor. Where T' is the time step, N is the number of nodes, C is the feature dimension, and d k This represents the total dimension of the vector.

[0074] In an optional implementation, in step S100, the spatiotemporal feature fusion further includes using the STL algorithm to separate the trend, periodic and residual terms of the logistics data, and using a graph attention network (GAT) to dynamically weight neighboring nodes based on the physical distance of the track. Each time component is then spliced ​​with the spatial embedding and input into a fully connected layer for fusion.

[0075] In another optional implementation, in step S100, the spatiotemporal feature fusion may further include constructing a spatiotemporal graph: nodes are workstations, edges are rail connections, edge weights are transportation time, and the STGCN model is used to synchronously process temporal convolution (capturing speed changes) and spatial graph convolution, with the output directly used as the fusion feature.

[0076] It should be noted that the trigonometric function embedding of this invention explicitly encodes the time order, avoiding information loss in time decomposition. The Laplace smoothing of GCN ensures spatial topological stability, making it more suitable for rigid semiconductor rail structures compared to GAT / STGCN.

[0077] In step S200, the encoded feature extraction process includes steps S201 to S203:

[0078] S201: Perform time-series dynamic awareness operations to capture temporal dynamic characteristics through local context feature extraction;

[0079] S202: Perform spatial relationship adaptive operation, dynamically generate a relationship strength matrix based on node features and adjust spatial connection weights;

[0080] S203: Perform feature stabilization on the processing results and enhance the feature representation capability through nonlinear feature enhancement.

[0081] Furthermore, in step S201, the local context feature extraction includes steps C1 to C4:

[0082] C1: Project the input features into query, key, and value representations;

[0083] C2: Perform sliding window feature extraction on the query matrix and key matrix, expressed by the formula:

[0084]

[0085] Where Q is the query matrix, K is the key matrix, V is the value matrix, and H is the total number of attention heads. h 1 For the output of the h-th attention head, W O To output the projection matrix, The operation involves concatenating the outputs of all heads. TrSelfAttention(Q, K, V) represents the temporal self-attention output, and TrAttention is the attention computation based on the convolution results. and is the convolution kernel corresponding to the h-th head.

[0086] C3: Calculate the attention weights and weighted fusion value features, expressed by the formula:

[0087]

[0088] Where Q is the query matrix, K is the key matrix, V is the value matrix, Attention is the scaled dot product attention, and d k Let QK be the total dimension of the vector, softmax be the normalization exponential function, and QK be the vector dimension. T For query-key similarity matrix.

[0089] C4: Performs multiple sets of processing in parallel and merges the results; the formula is as follows:

[0090]

[0091] Where Multi-HeadOutput is the merging of multiple head outputs, H is the total number of attention heads, and head h This is the output of the h-th attention head.

[0092] In this embodiment of the application, step S202, the execution of the spatial relationship adaptive operation includes steps D1 to D3:

[0093] D1: Based on node features, calculate the similarity measure between node features, expressed by the formula:

[0094]

[0095] Among them, S t The dynamic correlation matrix is... N is the number of nodes. This is the transpose of the node feature matrix. Let d be the node feature matrix. k This represents the total dimension of the vector.

[0096] D2: Dynamic correlation weights are generated using a normalized exponential function, expressed by the following formula:

[0097]

[0098] Where ⊙ represents element-wise multiplication, G is the static weight matrix, and W... (l) σ represents the weights of the l-th layer GCN, σ is the activation function, and DGCN is the dynamic graph convolution operation.

[0099] D3: Use the adjusted matrix to weight and aggregate the logistics status of neighboring nodes.

[0100] In an optional implementation, in step S202, the execution of the spatial relationship adaptation operation further includes pre-training an LSTM to predict the future positions of each OHT, generating a dynamic adjacency matrix (connecting if the physical distance is less than a threshold) based on the predicted positions, and using a static GCN to aggregate features.

[0101] In another alternative implementation, a reward function is defined, node states are encoded using a GNN, and the adjacency matrix is ​​output by an RL agent to adjust actions and update node connection strategies online.

[0102] It should be noted that the node feature similarity calculation responds to changes in congestion status in real time without the need for a pre-trained model, and the dot product operation directly couples physical guide rail constraints with dynamic logistics, making it more stable and reliable than the RL solution.

[0103] In step S300, the decoding prediction process includes steps S301 to S303:

[0104] S301: Utilizes a causal constraint attention mechanism and employs unidirectional information flow control to ensure temporal causality;

[0105] S302: Implement cross-sequence association modeling, establish the association between the decoded sequence and the encoded output, and calculate the cross-sequence attention formula as follows:

[0106]

[0107] Among them, head h For cross-sequence attention, For causal convolution operations, MaskTrAttention performs masked temporal attention computation, Q' is the decoder query matrix, K' is the decoder key matrix, and V' is the decoder value matrix. and The weights of the causal convolution kernel for the h-th head are... The values ​​are linear projection weights of the value matrix, * represents the standard 1D convolution operation, and TrAttention represents the temporal-aware attention computation.

[0108] S303: Combines spatial relationship adaptive operation to process spatial features and outputs multi-node, multi-time-step prediction sequences.

[0109] Furthermore, in this embodiment of the application, in step S301, the causal constraint attention mechanism includes performing causal convolution on the decoder input sequence and forcing causality through a mask matrix, prohibiting future time steps from participating in the current calculation, and each head focusing on different time period patterns, as expressed by the formula:

[0110]

[0111] Among them, head h ' is a mask-based time-series-aware multi-head self-attention mechanism. For causal convolution operations, MaskTrAttention performs masked temporal attention computation, Q' is the decoder query matrix, K' is the decoder key matrix, and V' is the decoder value matrix. and The weights of the causal convolution kernel for the h-th head are... The linear projection weights of the value matrix.

[0112] In an optional implementation, in step S301, the causal constraint attention mechanism further includes designing a linear state equation, taking the logistics state (speed / flow) as an observed variable, and using Kalman filtering for recursive prediction.

[0113] In another alternative implementation, in step S301, the causal constraint attention mechanism may further include training the Transformer with complete time-series data, using only historical data, forcing causality through attention masks, and constraining the student model to mimic the teacher model's output through distillation loss.

[0114] In this embodiment of the application, step S300, the determination of the blockage state, includes steps S311 to S313:

[0115] S311: Read the average speed of each node from the logistics status matrix sequence;

[0116] S312: Take the 10th percentile of historical speed as the threshold. When the average speed is lower than the preset threshold, it is determined that the current node is blocked.

[0117] S313: When the real-time speed is less than the threshold and continues for 3 time steps, a congestion alarm is triggered.

[0118] In an optional implementation, step S300, the congestion status determination further includes constructing a feature vector: [speed, track occupancy, queue length], training an SVM classifier, and triggering a warning if the congestion probability threshold is greater than 0.8.

[0119] In another optional implementation, step S300, the process of determining the blockage status may further include establishing a digital twin model of the factory in FlexSim, injecting real-time logistics data, simulating future states, and detecting OHT stagnation events in the simulation as blockages.

[0120] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0121] Example 3, referring to Figure 3 and Figure 4 This is the second embodiment of the present invention, which provides a method for predicting logistics blockages in semiconductor wafer manufacturing plants. To verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0122] Two sets of logistics datasets from an automated material handling simulation system in a semiconductor manufacturing smart factory and one set of logistics datasets from an automated material handling system in a semiconductor smart factory in Shanghai were used to verify the effectiveness of the proposed model that integrates Transformer and Dynamic Graph Convolutional Network (ITDGCN). The datasets contain real-time logistics data from multiple workstations and conveyor paths within the semiconductor smart factory, covering detailed information such as the flow rate, occupancy rate, and average speed index of the transport rails over a period of one month. Table 1 summarizes the information from the three datasets. Data was collected every 30 seconds, covering different production load scenarios, including peak and off-peak periods. All data were standardized (normalized to the [0,1] interval).

[0123] Table 1 Description of the three datasets

[0124]

[0125] The Transformer encoder and decoder have 4 layers, convolutional layers have 3 layers, attention heads have 8 layers, the model dimension is 64, the learning rate is 0.001, and the dropout probability is 0.2. The optimizer is Adam, the loss function is mean absolute error (MAE), the number of training epochs is 100, and the batch size is 4.

[0126] 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:

[0127]

[0128]

[0129] Where n is the number of samples, y i For the i-th actual observation, Let be the i-th predicted value.

[0130] To verify the effectiveness of the invented ITDGCN model, it was compared with other baseline models to predict logistics data for the next fifteen minutes. The experimental results were compared and analyzed using the three performance evaluation indicators mentioned above. The results are shown in Table 2. The ITDGCN method provided by this invention outperforms other methods in all three performance evaluation indicators.

[0131] Table 2 Performance comparison of the present invention with other methods

[0132]

[0133] To further validate the model's effectiveness, multi-step prediction experiments were conducted on the D3 dataset to compare it with other baseline models, predicting material flow rates for the next 5, 10, 15, and 20 minutes. The results are as follows: Figure 3 As shown, ITDGCN exhibits optimal or near-optimal performance. Generally, prediction accuracy decreases with increasing prediction time steps, leading to a performance decline for all models. However, ITDGCN shows the least performance degradation in most cases, demonstrating a more pronounced performance stability advantage as the prediction time steps increase, thus validating the model's good stability.

[0134] To more intuitively demonstrate the effectiveness of the ITDGCN model, a key guide rail node from the D3 dataset is selected to visualize the prediction results, and the predicted value is compared with the actual value. Figure 4 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.

[0135] Example 4, refer to Figure 5 This is the fourth embodiment of the present invention. This embodiment provides a semiconductor wafer manufacturing plant logistics congestion prediction system, including a GUI module, a spatiotemporal information input module, a Transformer encoder module, and a Transformer decoder module.

[0136] The GUI module is used to receive real-time logistics data from the semiconductor manufacturing smart factory collected by the user, including flow rate, occupancy rate and average speed index information, and use the data as raw data.

[0137] The spatiotemporal information input module is used to convert the collected logistics data into a high-dimensional representation containing spatiotemporal features, and enhance the model's ability to perceive spatiotemporal information through a spatiotemporal location embedding strategy.

[0138] The Transformer encoder module is used to capture the temporal and spatial dynamic features of logistics data through a time-aware multi-head self-attention layer and a dynamic graph convolutional layer.

[0139] The Transformer decoder module is used to predict the future multi-time step logistics state matrix sequence by using an autoregressive approach, combining masked time-aware multi-head self-attention and cross-sequence time-aware multi-head self-attention. It also determines whether a node is congested by using the corresponding speed exponent magnitude. Finally, the generated prediction data is transmitted to the GUI module for path planning.

[0140] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0141] Example 5, the fifth embodiment of the present invention, differs from the previous four embodiments in that:

[0142] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present 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 described in the various embodiments of the present 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.

[0143] 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.

[0144] More specific examples of computer-readable media (a non-exhaustive list) 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 the program can be printed, because the program 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.

[0145] 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.

Claims

1. A method for predicting logistics congestion in semiconductor wafer manufacturing plants, characterized by: include, Historical logistics data from a smart semiconductor manufacturing factory is collected and preprocessed. The preprocessed data is then fused with spatiotemporal features to generate high-dimensional representation data. Encoding feature extraction is performed on high-dimensional representation data. Spatiotemporal features are extracted through time-series dynamic perception and spatial relationship adaptive operations, and spatiotemporal feature fusion is performed. The processing results are then integrated with the complex spatiotemporal features of the encoded output through cross-sequence feature integration. The integrated features are decoded and predicted. The logistics state matrix sequence is predicted through causal constraint attention mechanism and cross-sequence association modeling. The congestion status is determined based on the logistics state matrix sequence.

2. The semiconductor wafer manufacturing plant logistics congestion prediction method as described in claim 1, characterized in that: The spatiotemporal feature fusion includes generating periodic position codes for the time dimension and extending them to the spatial dimension; Generate topological structure encoding for the spatial dimension, enhance spatial relationship representation through neighborhood information propagation, and extend the results to the temporal dimension; The encoding representations of the time and space dimensions are combined with multidimensional features.

3. The semiconductor wafer manufacturing plant logistics congestion prediction method as described in claim 2, characterized in that: The encoded feature extraction process includes performing a time-series dynamic perception operation to capture temporal dynamic characteristics through local context feature extraction. Perform spatial relationship adaptive operation, dynamically generate relationship strength matrix based on node features and adjust spatial connection weights; The processing results are then subjected to feature stabilization. Enhance feature representation capabilities through nonlinear feature enhancement.

4. The semiconductor wafer manufacturing plant logistics congestion prediction method as described in claim 3, characterized in that: The decoding prediction process includes using a causal constraint attention mechanism and employing unidirectional information flow control to ensure temporal causality; Implement cross-sequence association modeling to establish the association between the decoded sequence and the encoded output; Adaptive operations combining spatial relationships are used to process spatial features; Output the predicted sequence with multiple nodes and multiple time steps.

5. The semiconductor wafer manufacturing plant logistics congestion prediction method as described in claim 4, characterized in that: The generation of periodic position codes includes calculating trigonometric function position vectors based on time step indices, constructing a time-dimensional encoding matrix, and extending the encoding matrix along the spatial dimension. Constructing the time-dimensional encoding matrix includes the shape of the input data. For each time step index, the vector is calculated to generate a time dimension encoding matrix according to the formula, which is expressed as: Among them, E TP The embedding matrix is ​​a time position matrix, where t is the time step index and d is the dimension index. k The total dimension of the vector. Where T is the frequency adjustment factor, T' is the number of time steps, N is the number of nodes, and C is the feature dimension; Extending the encoding matrix along the spatial dimension includes E TP The copy is extended and replicated across all nodes to form an extended temporal location embedding tensor. The generated topology encoding includes generating an initial embedding matrix for each node, performing Laplacian smoothing through multi-layer GCN operations, aggregating neighboring node information, and each layer including adjacency matrix normalization and feature transformation operations. The formula for generating the initial embedding matrix is ​​expressed as: The formula for multi-layer GCN operation is expressed as follows: in, For the initial spatial embedding of node i, W s For trainable weights, b s For trainable bias, s i Let be the static feature vector of node i. Let A be the spatial embedding matrix of the l-th layer, and let I be the adjacency matrix. N It is the identity matrix. To add the adjacency matrix after adding self-loops, W is the degree matrix. (l) Here, σ represents the weights of the l-th layer GCN, and σ is the activation function. Let be the spatial embedding matrix of the (l-1)th layer.

6. The semiconductor wafer manufacturing plant logistics congestion prediction method as described in claim 5, characterized in that: The local context feature extraction includes projecting the input features into query, key, and value representations, performing sliding window feature extraction on the query and key, calculating attention weights and weighted fusing value features, and performing multiple sets of processing in parallel and merging the results. The relationship strength matrix includes calculating the similarity measure between node features, generating dynamic correlation weights through a normalized exponential function, and adjusting the static connection matrix and the dynamic weight matrix element by level. The cross-sequence feature integration includes applying unidirectional information flow control to the decoder query sequence, applying global context extraction to the encoder key sequence, calculating cross-sequence attention weights, and fusing features.

7. The semiconductor wafer manufacturing plant logistics congestion prediction method as described in claim 6, characterized in that: The process of determining the congestion status includes extracting the velocity index of each node from the predicted sequence and comparing the velocity index with dynamic threshold conditions to determine the congestion status.

8. A semiconductor wafer manufacturing plant logistics congestion prediction system, employing the semiconductor wafer manufacturing plant logistics congestion prediction method as described in any one of claims 1 to 7, characterized in that, It includes a GUI module, a spatiotemporal information input module, a Transformer encoder module, and a Transformer decoder module; The GUI module is used to receive real-time logistics data from the semiconductor manufacturing smart factory collected by the user, including flow rate, occupancy rate and average speed index information, and use the data as raw data. The spatiotemporal information input module is used to convert the collected logistics data into a high-dimensional representation containing spatiotemporal features, and enhance the model's ability to perceive spatiotemporal information through a spatiotemporal location embedding strategy. The Transformer encoder module is used to capture the temporal and spatial dynamic features of logistics data through a time-aware multi-head self-attention layer and a dynamic graph convolutional layer. The Transformer decoder module is used to predict the future multi-time step logistics state matrix sequence by using an autoregressive approach, combining masked time-aware multi-head self-attention and cross-sequence time-aware multi-head self-attention. It also determines whether a node is congested by using the corresponding speed exponent magnitude. Finally, the generated prediction data is transmitted to the GUI module for path planning.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the semiconductor wafer manufacturing plant logistics congestion prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the semiconductor wafer manufacturing plant logistics congestion prediction method as described in any one of claims 1 to 7.

Images