A semiconductor machine multi-interface wireless communication method based on edge computing

Abstract

The application discloses a kind of based on edge computing's semiconductor machine multi-interface wireless communication method, belong to semiconductor manufacturing and industrial internet of things technical field. Including the following steps: S1, in the control cabinet of each semiconductor machine, deploy industrial-grade edge computing node, edge computing node integrates multimode wireless communication module, protocol conversion module and AI acceleration module;S2, continuously collect multimodal historical communication data between semiconductor machine and host computer, multimodal historical communication data include electrical characteristic data, time series characteristic data, protocol characteristic data and environmental characteristic data. Construct multistage feature extraction architecture, design differentiating feature extraction operator for electrical, time series, protocol, environment four kinds of multimodal data, accurately capture the nonlinear dynamic change law of communication state. Compared with traditional CNN / LSTM single architecture, model communication parameter prediction accuracy is improved, generalization ability is significantly enhanced, can be adapted to different types, different process stages of various semiconductor machine.

Description

Technical Field

[0001] This invention relates to the fields of semiconductor manufacturing and industrial Internet of Things (IoT) technology, and in particular to a multi-interface wireless communication method for semiconductor equipment based on edge computing. Background Technology

[0002] In the semiconductor manufacturing industry, equipment communication systems are the core infrastructure of the production line. As semiconductor processes move to 3nm and below, a single EUV lithography machine generates more than 5GB of data per second, placing unprecedented demands on the bandwidth, latency, reliability, and security of communication systems.

[0003] In existing technologies, semiconductor equipment communication primarily uses wired Ethernet, such as the HSMS-SS protocol. However, wired communication has the following inherent drawbacks:

[0004] The deployment is complex, requires a lot of wiring, has a long construction period, high cost, and occupies valuable cleanroom space;

[0005] It lacks flexibility, and rewiring is required when adjusting production lines or relocating equipment, which affects production progress.

[0006] Maintenance is difficult, cable fault diagnosis is challenging, and rotary joints and connectors are prone to wear, leading to communication interruptions.

[0007] It has poor scalability, and the network architecture needs to be redesigned when adding new devices.

[0008] To address the challenges of wired communication, several wireless communication solutions have emerged in recent years. However, semiconductor manufacturing environments are extremely complex, containing numerous metal devices, precision instruments, and high-frequency electronic equipment, leading to severe electromagnetic interference and significant multipath effects. Traditional wireless communication technologies perform poorly in such environments, exhibiting issues such as high data transmission error rates, difficulty in guaranteeing real-time performance, and poor protocol compatibility.

[0009] Existing deep learning-based communication optimization methods mostly employ a single neural network architecture, such as CNN or LSTM, which cannot fully exploit the complex features in multimodal communication data. Especially for application scenarios like semiconductor machine communication, which are characterized by strong temporality, multimodality, and strong interference, traditional feature extraction methods struggle to accurately capture the dynamic changes in communication states, resulting in insufficient model prediction accuracy and generalization ability.

[0010] Therefore, there is an urgent need to develop a multi-interface wireless communication method for semiconductor equipment based on edge computing with stronger feature extraction capabilities, which can achieve highly reliable and low-latency wireless communication in complex electromagnetic environments. Summary of the Invention

[0011] Purpose of the invention: The purpose of this invention is to provide a multi-interface wireless communication method for semiconductor equipment based on edge computing; it can solve the problems of low reliability, poor real-time performance, and insufficient feature extraction capability of existing wireless communication technologies in the complex electromagnetic environment of semiconductor plants.

[0012] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a multi-interface wireless communication method for semiconductor equipment based on edge computing, comprising the following steps:

[0013] S1. Deploy industrial-grade edge computing nodes in the control cabinet of each semiconductor machine. The edge computing nodes integrate multi-mode wireless communication modules, protocol conversion modules and AI acceleration modules.

[0014] S2. Continuously collect multimodal historical communication data between the semiconductor equipment and the host computer. The multimodal historical communication data includes electrical characteristic data, timing characteristic data, protocol characteristic data and environmental characteristic data.

[0015] S3. Perform multi-stage preprocessing operations on the multimodal historical communication data, including data synchronization, data cleaning, data standardization, and data augmentation, to obtain a standardized training dataset.

[0016] S4. Construct a wireless communication optimization model based on a multi-level feature extraction architecture of first-level features, second-level features, and third-level features, and train the model end-to-end using a standardized training dataset.

[0017] S5. Deploy the trained wireless communication optimization model to the AI ​​acceleration module of the edge computing node and perform INT8 quantization and TensorRT inference optimization.

[0018] S6. The edge computing node collects the current multimodal communication status data in real time, performs the same preprocessing operation as in step S3 on the real-time data, and then inputs it into the deployed wireless communication optimization model.

[0019] S7. The wireless communication optimization model sequentially performs the first-level feature extraction, the second-level feature extraction, and the third-level feature extraction, and outputs the optimal communication parameter configuration vector.

[0020] S8. The edge computing node sends control commands to the multi-mode wireless communication module through the PCIe 4.0 interface. It dynamically adjusts all the working parameters of the multi-mode wireless communication module according to the optimal communication parameter configuration vector to achieve low-latency and high-reliability wireless communication between the semiconductor machine and the host computer.

[0021] S9. Edge computing nodes periodically upload newly collected communication data to the cloud server, use the new data to incrementally train the model, and then send the updated model parameters back to the edge computing nodes.

[0022] Furthermore, in step S1, the multi-mode wireless communication module simultaneously supports at least four wireless communication protocols, and the protocol conversion module supports hardware-accelerated conversion of at least six industrial protocols.

[0023] Furthermore, in step S2, the multimodal historical communication data includes: electrical characteristic data, timing characteristic data, protocol characteristic data, and environmental characteristic data;

[0024] Electrical characteristic data include: signal strength, signal-to-noise ratio, bit error rate, channel quality indication, received signal success rate, and interference signal power;

[0025] Timing characteristic data includes: transmission delay, jitter, packet arrival time interval, number of retransmissions, number of timeouts, and frame interval.

[0026] Protocol characteristic data includes: protocol type, packet size, frame type, sequence number, checksum, and control fields;

[0027] Environmental characteristic data include: equipment temperature, equipment humidity, electromagnetic interference intensity, vibration intensity, air pressure, and cleanliness level.

[0028] Furthermore, the specific steps of the multi-stage preprocessing in step S3 are as follows:

[0029] S31. High-precision synchronization of multimodal data: Based on the IEEE 1588 precise time protocol, nanosecond-level synchronization of different sensors and data sources is achieved, generating a multimodal data sequence with a unified timestamp;

[0030] S32. Multi-dimensional data cleaning: The isolated forest algorithm is used to identify and remove outliers, and the sliding window midpoint filter is used to remove high-frequency noise.

[0031] S33. Multimodal data differentiation standardization: For numerical data, Z-score standardization is used to convert it into a standard normal distribution with a mean of 0 and a variance of 1; for categorical data, one-hot coding is used; for time series data, min-max standardization is used to scale the data to the [0,1] interval.

[0032] S34. Intelligent Enhancement of Multimodal Data: Conditional Generative Adversarial Networks are used to generate synthetic data, expand the training dataset, and label the synthetic data.

[0033] Furthermore, in step S4, the extraction process of the first-level features is specifically as follows:

[0034] S411, Branching: The preprocessed multimodal input data is divided into four fully parallel feature extraction branches, namely the electrical feature branch, the timing feature branch, the protocol feature branch, and the environmental feature branch;

[0035] S412, Multi-scale Convolution Operation: Parallel convolution operations are performed on each branch using three sets of one-dimensional convolution kernels of different scales.

[0036] Electrical feature branches: convolution kernel sizes of 3, 5, and 7, stride of 1, and padding method of the same;

[0037] Temporal feature branches: convolutional kernel sizes are 5, 9, and 13, stride is 1, and padding method is the same;

[0038] Protocol feature branches: convolutional kernel sizes are 2, 4, and 6, stride is 1, and padding method is the same;

[0039] Environmental feature branch: Convolutional kernel sizes are 7, 11, and 15, stride is 1, and padding method is the same;

[0040] S413, Batch Normalization and Activation: Batch normalization is performed on the three sets of convolution results for each branch, and then a nonlinear transformation is performed using the LeakyReLU activation function with a negative slope of 0.1;

[0041] S414, Multi-Dilution Rate Dilated Convolution: Three sets of dilated convolution operations with different dilation rates are applied to the activation results of each branch to expand the receptive field.

[0042] Electrical characteristic branches: void ratios of 1, 2, and 4 respectively;

[0043] Temporal feature branches: void ratios of 2, 4, and 8, respectively;

[0044] Protocol characteristic branches: void ratios of 1, 3, and 5 respectively;

[0045] Environmental characteristic branches: void ratios of 4, 8, and 16% respectively;

[0046] S415, Intra-branch feature fusion: The three sets of dilated convolution results of each branch are concatenated along the channel dimension to obtain the fused features of each branch;

[0047] S416. Cross-branch feature concatenation and pooling: Concatenate the fused features of the four branches along the channel dimension to obtain the first-level fused features; perform max pooling operation with a window size of 2 and a step size of 2 on the first-level fused features to obtain the first-level features.

[0048] Furthermore, in step S4, the extraction process for the second-level features is specifically as follows:

[0049] S421, Bidirectional Gated Recurrent Unit Processing: The first-level features are input into a timing processing module consisting of two bidirectional gated recurrent units connected in series, and the forward and backward timing-dependent features are extracted respectively.

[0050] S422. Bidirectional feature concatenation: The forward and backward features output by each Bi-GRU are concatenated along the channel dimension to obtain bidirectional temporal features.

[0051] S423, Temporal Attention Weighting: Introducing a temporal attention mechanism to calculate the importance weight of each feature at each time step and weighting the bidirectional temporal features;

[0052] S424, Spatial Attention Weighting: Introducing a spatial attention mechanism to calculate the importance weight of each feature dimension, and then further weighting the features after temporal attention weighting;

[0053] S425, Residual Connection and Layer Normalization: Perform residual connection between the spatiotemporal attention-weighted features and the original bidirectional temporal features, and then perform layer normalization.

[0054] S426. Feature Transformation: The normalized features are transformed nonlinearly through a fully connected layer to obtain the second-level features.

[0055] Furthermore, in step S4, the extraction process of the third-level features is specifically as follows:

[0056] S431. Construction of graph structure data: Map the second-level features to undirected graph structure data G=(V,E), where each node in the node set V corresponds to a feature dimension; the weights of the edge set E are determined by the Pearson correlation coefficient between features, and only edges with an absolute value of the correlation coefficient greater than a preset value are retained.

[0057] S432, Multi-layer graph convolution operation: A two-layer graph convolutional neural network is used to perform convolution operations on graph structure data to extract local correlation features and global topological features between nodes;

[0058] S433, Cross-modal attention fusion: Introducing a cross-modal attention mechanism, calculating the correlation weights between four modal features (electrical, timing, protocol, and environment), and adaptively weighting and fusing features from different modalities based on the correlation weights to obtain cross-modal fused features;

[0059] S434. Contrastive learning for anomaly perception: Construct an anomaly perception sub-network based on contrastive learning, using normal communication state features as positive samples and abnormal communication state features as negative samples. Learn the feature differences between normal and abnormal states through a contrastive loss function to obtain anomaly perception features.

[0060] S435, Multi-feature integrated stitching: The graph convolution features, cross-modal fusion features and anomaly perception features are stitched together in the channel dimension to obtain integrated features;

[0061] S436, Multilayer Perceptron Transformation: The comprehensive features are transformed nonlinearly through two fully connected layers, with ReLU activation function for both, to obtain the third-level features.

[0062] Furthermore, in step S4, the wireless communication optimization model training process is specifically as follows:

[0063] S441. Dataset partitioning: The standardized training dataset is randomly divided into training set, validation set and test set according to a preset ratio, and the partitioning process maintains a consistent distribution of data in each category.

[0064] S442. Transfer learning pre-training: Use publicly available industrial communication datasets to pre-train the model, set the number of pre-training rounds and the learning rate, and initialize the model parameters;

[0065] S443, Optimizer and Learning Rate Settings: The AdamW optimizer is used, and the initial learning rate and weight decay coefficient are set to 0. The cosine annealing learning rate scheduling strategy is adopted.

[0066] S444, Multi-task joint loss function: Training is performed using a multi-task joint loss function. The total loss is the weighted sum of the mean squared error loss of communication parameter prediction, the binary cross-entropy loss of anomaly detection, and the cross-entropy loss of protocol classification.

[0067] S445, Mixed Precision Training: Employs FP16 mixed precision training technology to accelerate the model training process and reduce GPU memory usage;

[0068] S446. Early stopping mechanism: When the validation set loss no longer decreases after a preset number of consecutive rounds, the early stopping mechanism is triggered to stop training and save the current optimal model parameters.

[0069] S447. Model Performance Evaluation: The trained model is evaluated using a test set. Evaluation metrics include mean absolute error, root mean square error, accuracy, precision, recall, and F1 score. The model is considered qualified when all evaluation metrics meet the preset standards.

[0070] Furthermore, in step S7, the optimal communication parameter configuration vector includes: wireless communication protocol selection, channel switching, transmit power adjustment, transmission rate adjustment, data packet size adjustment, retransmission count adjustment, modulation method adjustment, coding method adjustment, and time slot allocation adjustment.

[0071] Furthermore, in step S9, the specific process of incremental training and model update is as follows:

[0072] S91. Federated Learning Cluster Construction: Form a federated learning cluster by combining all edge computing nodes in the same workshop and designate the cloud server as the aggregation node;

[0073] S92, Local Model Training: Each edge computing node uses locally collected new communication data to incrementally train the model;

[0074] S93, Encrypted Gradient Upload: Each edge computing node uses homomorphic encryption technology to encrypt the local model gradient, and uploads the encrypted gradient to the cloud server once a day;

[0075] S94, Global Gradient Aggregation: The cloud server uses a federated averaging algorithm to aggregate the encrypted gradients uploaded by all edge computing nodes to obtain the global model gradient;

[0076] S95, Global Model Update: The cloud server updates the global model parameters based on the global model gradient and distributes the updated global model parameters to all edge computing nodes;

[0077] S96. Local Model Hot Update: Edge computing nodes can update their local models using hot update technology without interrupting normal communication services, thus achieving continuous model optimization.

[0078] Beneficial effects: It replaces the traditional HSMS-SS wired Ethernet solution, eliminating the need for large-scale cabling, reducing production line deployment cycle, and saving valuable cabling space in cleanrooms; it eliminates the need for rewiring when adjusting production lines or relocating equipment, shortening equipment uptime; it avoids occasional communication interruptions caused by cable wear and connector oxidation, reducing maintenance costs; it supports plug-and-play devices, and the network architecture can be rapidly expanded without replanning.

[0079] By dynamically switching protocols and channels through a multi-mode wireless communication module, combined with AI-driven real-time parameter adaptive optimization, the data transmission error rate is reduced in extreme environments with strong electromagnetic interference and significant multipath effects in semiconductor plants. The end-to-end transmission latency fully meets the high bandwidth and hard real-time data transmission requirements of core equipment such as EUV lithography machines at 3nm and below process nodes, which are above 5GB / s.

[0080] A multi-level feature extraction architecture is constructed, and differentiated feature extraction operators are designed for four types of multimodal data: electrical, timing, protocol, and environmental, to accurately capture the nonlinear dynamic changes in communication states. Compared with the traditional single CNN / LSTM architecture, the model's accuracy in predicting communication parameters is improved, its generalization ability is significantly enhanced, and it can be adapted to various semiconductor equipment of different models and process stages.

[0081] The trained model is deployed on a dedicated AI acceleration module on an edge computing node, and the inference speed is greatly improved through INT8 quantization and TensorRT inference optimization, effectively reducing the latency of single-frame data processing and meeting the hard real-time requirements of industrial control. All real-time computing and control logic is completed at the edge, without relying on cloud servers, avoiding control latency and data loss caused by wide area network congestion.

[0082] The protocol conversion module supports hardware-accelerated conversion of at least six mainstream industrial protocols, such as HSMS-SS, SECS / GEM, and ModbusTCP. It does not require modification of the native communication interface of existing semiconductor equipment and can directly connect to new and old equipment to achieve a smooth upgrade of the communication system. The multi-mode wireless module simultaneously supports at least four wireless protocols, such as Wi-Fi6E and 5GNR-U, and can dynamically select the optimal communication method according to service priority.

[0083] Employing a federated learning distributed training architecture, edge nodes within the same workshop form a cluster for incremental training. The original process data is stored locally throughout the process, with only the homomorphically encrypted model gradients being uploaded. This not only strictly protects the security of core process data in semiconductor manufacturing but also leverages the massive amounts of data from multiple machines to improve the model's global generalization ability. It supports hot model updates, enabling model iteration without interrupting production and continuously optimizing communication performance. Attached Figure Description

[0084] Figure 1 This is a flowchart illustrating the method. Detailed Implementation

[0085] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0086] Example 1

[0087] Industrial-grade edge computing node deployment

[0088] An industrial-grade edge computing node is deployed within a semiconductor equipment control cabinet. This node integrates the following core modules:

[0089] Multi-mode wireless communication module: Simultaneously supports four wireless communication protocols: Wi-Fi 6E, 5G NR-U, Bluetooth 5.3, and Zigbee 3.0, and is equipped with a high-gain antenna array.

[0090] Protocol Conversion Module: Implements conversion of six industrial protocols—HSMS-SS, SECS / GEM, ModbusTCP, EtherNet / IP, PROFINET, and OPCUA—based on FPGA hardware acceleration.

[0091] AI acceleration module: Equipped with a dedicated AI inference chip, supporting INT8 quantization and TensorRT acceleration.

[0092] Data acquisition module: Connects to the communication interface and sensors of the semiconductor equipment via a PCIe 4.0 interface.

[0093] Multimodal historical communication data acquisition

[0094] Continuously collect 30 days of multimodal historical communication data between the semiconductor equipment and the host computer at a frequency of 1kHz. The specific data types are as follows:

[0095] Electrical characteristic data: signal strength, signal-to-noise ratio, bit error rate, channel quality indication, received signal success rate, and interference signal power.

[0096] Timing characteristics data: transmission delay, jitter, data packet arrival time interval, number of retransmissions, number of timeouts, and frame interval.

[0097] Protocol characteristic data: protocol type, packet size, frame type, sequence number, checksum, and control fields.

[0098] Environmental characteristics data: equipment temperature, equipment humidity, electromagnetic interference intensity, vibration intensity, air pressure, and cleanliness level.

[0099] Multi-stage preprocessing operations

[0100] The following preprocessing steps are performed sequentially on the acquired raw multimodal data:

[0101] High-precision synchronization of multimodal data: Based on the IEEE 1588 precision time protocol, nanosecond-level time synchronization is achieved for all sensors and data sources. A unified timestamp is added to each data point to generate a time-aligned multimodal data sequence.

[0102] Multi-dimensional data cleaning: The Isolation Forest algorithm was used to identify and remove outliers, with an outlier threshold set to 0.8; a sliding window mid-range filter with a window size of 15 was used to remove high-frequency noise.

[0103] Multimodal data differentiation and standardization:

[0104] Numerical data (such as signal strength and transmission delay) are normalized using the Z-score method to be converted into a standard normal distribution with a mean of 0 and a variance of 1.

[0105] One-hot coding is used for categorical data (such as protocol type and frame type).

[0106] The time series data is scaled to the [0,1] interval using the min-max normalization method.

[0107] Multimodal data intelligent enhancement: Synthetic data is generated using conditional generative adversarial networks (CGAN) to expand the training dataset to 1.2 times the size of the original data. Professional engineers then annotate the synthetic data to obtain a standardized training dataset.

[0108] Wireless communication optimization model construction and training

[0109] A wireless communication optimization model based on a three-level feature extraction architecture is constructed and trained end-to-end using a standardized training dataset:

[0110] First-level feature extraction

[0111] The preprocessed multimodal input data is divided into four fully parallel feature extraction branches: electrical feature branch, temporal feature branch, protocol feature branch, and environmental feature branch.

[0112] For each branch, three sets of one-dimensional convolution kernels of different scales are used for parallel convolution operations, all with a stride of 1 and the same padding method:

[0113] Electrical feature branches: kernel sizes of 3, 5, and 7;

[0114] Temporal feature branches: convolution kernel sizes are 5, 9, and 13 respectively;

[0115] Protocol feature branches: convolutional kernel sizes of 2, 4, and 6;

[0116] Environmental feature branches: convolution kernel sizes are 7, 11, and 15 respectively.

[0117] Batch normalization was performed on the three sets of convolution results for each branch, and then a nonlinear transformation was performed using the LeakyReLU activation function with a negative slope of 0.1.

[0118] For the activation results of each branch, three sets of dilated convolution operations with different dilation rates are applied to expand the receptive field:

[0119] Electrical characteristic branches: void ratios of 1, 2, and 4 respectively;

[0120] Temporal feature branches: void ratios of 2, 4, and 8, respectively;

[0121] Protocol characteristic branches: void ratios of 1, 3, and 5 respectively;

[0122] Environmental characteristic branches: void ratios of 4, 8, and 16, respectively.

[0123] The three sets of dilated convolution results from each branch are concatenated along the channel dimension to obtain the fused features of each branch.

[0124] The fused features of the four branches are concatenated along the channel dimension to obtain the first-level fused features; a max pooling operation with a window size of 2 and a step size of 2 is performed on the first-level fused features to obtain the first-level features.

[0125] Second-level feature extraction

[0126] The first-level features are input into a time-series processing module consisting of two bidirectional gated recurrent units (Bi-GRU) connected in series. Each Bi-GRU has 128 hidden layer units and a dropout rate of 0.2, and forward and backward temporal dependency features are extracted respectively.

[0127] The forward and backward features output by each Bi-GRU are concatenated along the channel dimension to obtain bidirectional temporal features.

[0128] A time attention mechanism is introduced to calculate the importance weight of each feature at each time step, and to weight the bidirectional time series features.

[0129] A spatial attention mechanism is introduced to calculate the importance weight of each feature dimension, and the features after temporal attention weighting are weighted again.

[0130] The spatiotemporal attention-weighted features are residually connected with the original bidirectional temporal features, and then layer normalization is performed.

[0131] The normalized features are transformed nonlinearly by a fully connected layer to obtain the second-level features.

[0132] Third-level feature extraction

[0133] The second-level features are mapped to undirected graph structure data G=(V,E), where each node in the node set V corresponds to a feature dimension; the weights of the edge set E are determined by the Pearson correlation coefficient between the features, and only edges with an absolute value of correlation coefficient greater than 0.5 are retained.

[0134] A two-layer graph convolutional neural network is used to perform convolution operations on graph structure data to extract local correlation features and global topological features between nodes.

[0135] A cross-modal attention mechanism is introduced to calculate the correlation weights between four modal features: electrical, timing, protocol, and environment. Based on the correlation weights, the features of different modalities are adaptively weighted and fused to obtain cross-modal fused features.

[0136] An anomaly perception subnetwork based on contrastive learning is constructed, with normal communication state features as positive samples and abnormal communication state features as negative samples. The feature differences between normal and abnormal states are learned through a contrastive loss function to obtain anomaly perception features.

[0137] The graph convolutional features, cross-modal fusion features, and anomaly perception features are concatenated along the channel dimension to obtain a comprehensive feature.

[0138] The comprehensive features are nonlinearly transformed by two fully connected layers, with ReLU activation function for both, to obtain the third-level features.

[0139] Model training process

[0140] The standardized training dataset was randomly divided into training, validation, and test sets in a ratio of 7:2:1, while maintaining a consistent distribution of data across all categories during the partitioning process.

[0141] The model is pre-trained using publicly available industrial communication datasets to initialize model parameters.

[0142] The AdamW optimizer was used, with an initial learning rate of 1e-4 and a weight decay coefficient of 1e-5. A cosine annealing learning rate scheduling strategy was adopted.

[0143] Training is performed using a multi-task joint loss function, and the total loss is a weighted sum of the mean squared error loss for communication parameter prediction, the binary cross-entropy loss for anomaly detection, and the cross-entropy loss for protocol classification.

[0144] The FP16 mixed precision training technology is used to accelerate the model training process and reduce memory usage.

[0145] Set an early stopping mechanism. When the validation set loss no longer decreases for 10 consecutive rounds, trigger the early stopping mechanism, stop training, and save the current optimal model parameters.

[0146] The trained model is evaluated using a test set. Evaluation metrics include mean absolute error, root mean square error, accuracy, precision, recall, and F1 score. The model is considered qualified when all evaluation metrics meet the preset standards.

[0147] Model Deployment and Inference Optimization

[0148] The trained wireless communication optimization model is deployed to the AI ​​acceleration module of the edge computing node to perform INT8 quantization and TensorRT inference optimization, thereby improving the model inference speed and reducing resource consumption.

[0149] Real-time communication optimization

[0150] The edge computing node collects current multimodal communication status data in real time at a frequency of 1kHz, and performs the same preprocessing operation as in step 3 on the real-time data.

[0151] The preprocessed real-time data is then input into the deployed wireless communication optimization model.

[0152] The model sequentially performs the first, second, and third level feature extractions and outputs the optimal communication parameter configuration vector, including: wireless communication protocol selection, channel switching, transmit power adjustment, transmission rate adjustment, data packet size adjustment, retransmission count adjustment, modulation method adjustment, coding method adjustment, and time slot allocation adjustment.

[0153] Edge computing nodes send control commands to the multimode wireless communication module via the PCIe 4.0 interface, dynamically adjusting all operating parameters of the multimode wireless communication module according to the optimal communication parameter configuration vector, thereby achieving low-latency and high-reliability wireless communication between the semiconductor equipment and the host computer.

[0154] Incremental training and model update

[0155] All edge computing nodes within the same workshop are grouped into a federated learning cluster, with the cloud server designated as the aggregation node.

[0156] Each edge computing node uses locally collected new communication data to incrementally train the model.

[0157] Each edge computing node uses homomorphic encryption technology to encrypt the local model gradient, and uploads the encrypted gradient to the cloud server once a day.

[0158] The cloud server uses a federated averaging algorithm to aggregate the encrypted gradients uploaded by all edge computing nodes to obtain the global model gradient.

[0159] The cloud server updates the global model parameters based on the global model gradient and then distributes the updated global model parameters to all edge computing nodes.

[0160] Edge computing nodes can update their local models using hot update technology without interrupting normal communication services, thus enabling continuous model optimization.

[0161] Example 2

[0162] Cluster hardware deployment

[0163] The same industrial-grade edge computing nodes with the same configuration as in Example 1 are deployed in the control cabinet of each semiconductor machine in the cluster. All edge nodes form a wireless self-organizing network through multi-mode wireless communication modules. The cloud server is deployed in the factory data center and is responsible for federated learning gradient aggregation and global model management.

[0164] Distributed data acquisition and preprocessing

[0165] Each edge computing node continuously collects multimodal communication data from local semiconductor equipment for 15 days, with the data type being completely consistent with that in Example 1.

[0166] Each edge computing node independently performs multi-stage preprocessing operations locally, including IEEE 1588 data synchronization, isolated forest + sliding window mid-range filtering data cleaning, differential data standardization, and CGAN data augmentation.

[0167] All raw data and preprocessed standardized data are stored on local edge computing nodes and are not uploaded to cloud servers, ensuring the security of semiconductor manufacturing process data.

[0168] Federated learning distributed model training

[0169] The cloud server pre-trains an initial global model using a publicly available industrial communication dataset. The model architecture is exactly the same as in Example 1, and includes a three-level feature extraction module.

[0170] The cloud server distributes the initial global model parameters to all edge computing nodes within the cluster.

[0171] Each edge computing node uses a local standardized dataset to train the initial model locally, and the training process is the same as in Example 1.

[0172] After each round of local training is completed, the edge computing nodes use CKKS homomorphic encryption technology to encrypt the local model gradients before uploading them to the cloud server.

[0173] The cloud server uses a federated averaging algorithm to aggregate the encrypted gradients uploaded by all edge computing nodes to obtain the global model gradient.

[0174] The cloud server updates the global model parameters based on the global model gradient and then distributes the updated global model parameters to all edge computing nodes.

[0175] Repeat the above process of local training-gradient uploading-global aggregation-model distribution for a total of 15 rounds of federated iterations. Training stops when the average accuracy of the global model on the local validation set of all nodes reaches the preset standard.

[0176] Cluster collaborative communication optimization

[0177] All edge computing nodes deploy the trained global model and perform INT8 quantization and TensorRT inference optimization.

[0178] The best-performing edge computing node in the cluster is elected as the cluster coordinator, responsible for coordinating and allocating global channel resources.

[0179] Each edge computing node collects local multimodal communication status data in real time, inputs it into the local model to obtain the optimal communication parameter configuration vector, and reports the channel usage and interference status of the node to the cluster coordinator.

[0180] The cluster coordinator dynamically allocates global channel resources based on the communication status and optimal parameter requirements of all nodes, avoiding co-channel interference between multiple machines.

[0181] When a semiconductor machine experiences sudden electromagnetic interference that degrades communication quality, the local model automatically adjusts the optimal communication parameter configuration vector and switches to a wireless protocol and channel with less interference to ensure uninterrupted communication.

[0182] Continuous incremental updates

[0183] The federated learning cluster conducts an incremental training round at 2:00 AM every day. Each edge node uses the new communication data collected that day to conduct local training and uploads the encrypted gradient to the cloud.

[0184] The cloud server aggregates global gradients and updates the global model, then distributes the updates to all edge nodes.

[0185] Edge nodes use hot update technology to update the local model without affecting normal production communication.

[0186] A full model performance evaluation is conducted monthly, and the weight coefficients of the multi-task joint loss function are dynamically adjusted based on the evaluation results to continuously optimize model performance.

[0187] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A semiconductor machine multi-interface wireless communication method based on edge computing, characterized in that, Includes the following steps: S1. Deploy industrial-grade edge computing nodes in the control cabinet of each semiconductor machine. The edge computing nodes integrate multi-mode wireless communication modules, protocol conversion modules and AI acceleration modules. S2. Continuously collect multimodal historical communication data between the semiconductor equipment and the host computer. The multimodal historical communication data includes electrical characteristic data, timing characteristic data, protocol characteristic data and environmental characteristic data. S3. Perform multi-stage preprocessing operations on the multimodal historical communication data, including data synchronization, data cleaning, data standardization, and data augmentation, to obtain a standardized training dataset. S4. Construct a wireless communication optimization model based on a multi-level feature extraction architecture of first-level features, second-level features, and third-level features, and train the model end-to-end using a standardized training dataset. S5. Deploy the trained wireless communication optimization model to the AI ​​acceleration module of the edge computing node and perform INT8 quantization and TensorRT inference optimization. S6. The edge computing node collects the current multimodal communication status data in real time, performs the same preprocessing operation as in step S3 on the real-time data, and then inputs it into the deployed wireless communication optimization model. S7. The wireless communication optimization model sequentially performs the first-level feature extraction, the second-level feature extraction, and the third-level feature extraction, and outputs the optimal communication parameter configuration vector. S8. The edge computing node sends control commands to the multi-mode wireless communication module through the PCIe 4.0 interface. It dynamically adjusts all the working parameters of the multi-mode wireless communication module according to the optimal communication parameter configuration vector to achieve low-latency and high-reliability wireless communication between the semiconductor machine and the host computer. S9. Edge computing nodes periodically upload newly collected communication data to the cloud server, use the new data to incrementally train the model, and then send the updated model parameters back to the edge computing nodes.

2. The method for multi-interface wireless communication of semiconductor equipment based on edge computing according to claim 1, characterized in that: In step S1, the multi-mode wireless communication module simultaneously supports at least four wireless communication protocols, and the protocol conversion module supports hardware-accelerated conversion of at least six industrial protocols.

3. The method for multi-interface wireless communication of semiconductor equipment based on edge computing according to claim 1, characterized in that: In step S2, the multimodal historical communication data includes: electrical characteristic data, timing characteristic data, protocol characteristic data, and environmental characteristic data; Electrical characteristic data include: signal strength, signal-to-noise ratio, bit error rate, channel quality indication, received signal success rate, and interference signal power; Timing characteristic data includes: transmission delay, jitter, packet arrival time interval, number of retransmissions, number of timeouts, and frame interval. Protocol characteristic data includes: protocol type, packet size, frame type, sequence number, checksum, and control fields; Environmental characteristic data include: equipment temperature, equipment humidity, electromagnetic interference intensity, vibration intensity, air pressure, and cleanliness level.

4. The method of claim 1, wherein the method further comprises: receiving a request from the first wireless communication device; and transmitting a response to the request to the first wireless communication device. In step S3, the specific steps of the multi-stage preprocessing are as follows: S31. High-precision synchronization of multimodal data: Based on the IEEE 1588 precise time protocol, nanosecond-level synchronization of different sensors and data sources is achieved, generating a multimodal data sequence with a unified timestamp; S32. Multi-dimensional data cleaning: The isolated forest algorithm is used to identify and remove outliers, and the sliding window midpoint filter is used to remove high-frequency noise. S33. Multimodal data differentiation standardization: For numerical data, Z-score standardization is used to convert it into a standard normal distribution with a mean of 0 and a variance of 1; for categorical data, one-hot coding is used; for time series data, min-max standardization is used to scale the data to the [0,1] interval. S34. Intelligent Enhancement of Multimodal Data: Conditional Generative Adversarial Networks are used to generate synthetic data, expand the training dataset, and label the synthetic data.

5. The method of claim 1, wherein the method further comprises: receiving a request from the first wireless communication device; and transmitting a response to the request to the first wireless communication device. In step S4, the extraction process of the first-level features is specifically as follows: S411, Branching: The preprocessed multimodal input data is divided into four fully parallel feature extraction branches, namely the electrical feature branch, the timing feature branch, the protocol feature branch, and the environmental feature branch; S412, Multi-scale Convolution Operation: Parallel convolution operations are performed on each branch using three sets of one-dimensional convolution kernels of different scales. Electrical feature branches: convolution kernel sizes of 3, 5, and 7, stride of 1, and padding method of the same; Temporal feature branch: Convolutional kernel sizes are 5, 9, and 13, stride is 1, and padding method is the same; Protocol feature branches: convolutional kernel sizes are 2, 4, and 6, stride is 1, and padding method is the same; Environmental feature branch: Convolutional kernel sizes are 7, 11, and 15, stride is 1, and padding method is the same; S413, Batch Normalization and Activation: Batch normalization is performed on the three sets of convolution results for each branch, and then a nonlinear transformation is performed using the LeakyReLU activation function with a negative slope of 0.1; S414, Multi-Dilution Rate Dilated Convolution: Three sets of dilated convolution operations with different dilation rates are applied to the activation results of each branch to expand the receptive field. Electrical characteristic branches: void ratios of 1, 2, and 4 respectively; Temporal feature branches: void ratios of 2, 4, and 8, respectively; Protocol characteristic branches: void ratios of 1, 3, and 5 respectively; Environmental characteristic branches: void ratios of 4, 8, and 16% respectively; S415, Intra-branch feature fusion: The three sets of dilated convolution results of each branch are concatenated along the channel dimension to obtain the fused features of each branch; S416. Cross-branch feature concatenation and pooling: Concatenate the fused features of the four branches along the channel dimension to obtain the first-level fused features; perform max pooling operation with a window size of 2 and a step size of 2 on the first-level fused features to obtain the first-level features.

6. The method of claim 1, wherein the method further comprises: receiving a request from the first wireless communication device; and transmitting a response to the request to the first wireless communication device. In step S4, the extraction process for the second-level features is as follows: S421, Bidirectional Gated Recurrent Unit Processing: The first-level features are input into a timing processing module consisting of two bidirectional gated recurrent units connected in series, and the forward and backward timing-dependent features are extracted respectively. S422, Bidirectional Feature Concatenation: The forward and backward features output by each Bi-GRU are concatenated along the channel dimension to obtain bidirectional temporal features; S423, Temporal Attention Weighting: Introducing a temporal attention mechanism to calculate the importance weight of each feature at each time step and weighting the bidirectional temporal features; S424, Spatial Attention Weighting: Introducing a spatial attention mechanism to calculate the importance weight of each feature dimension, and then further weighting the features after temporal attention weighting; S425, Residual Connection and Layer Normalization: Perform residual connection between the spatiotemporal attention-weighted features and the original bidirectional temporal features, and then perform layer normalization. S426. Feature Transformation: The normalized features are transformed nonlinearly through a fully connected layer to obtain the second-level features.

7. The method of claim 1, wherein the method further comprises: receiving a request from the first wireless communication device; and transmitting a response to the request to the first wireless communication device. In step S4, the extraction process of the third-level features is specifically as follows: S431. Construction of graph structure data: Map the second-level features to undirected graph structure data G=(V,E), where each node in the node set V corresponds to a feature dimension; the weights of the edge set E are determined by the Pearson correlation coefficient between features, and only edges with an absolute value of the correlation coefficient greater than a preset value are retained. S432, Multi-layer graph convolution operation: A two-layer graph convolutional neural network is used to perform convolution operations on graph structure data to extract local correlation features and global topological features between nodes; S433, Cross-modal attention fusion: Introducing a cross-modal attention mechanism, calculating the correlation weights between four modal features (electrical, timing, protocol, and environment), and adaptively weighting and fusing features from different modalities based on the correlation weights to obtain cross-modal fused features; S434. Contrastive learning for anomaly perception: Construct an anomaly perception sub-network based on contrastive learning, using normal communication state features as positive samples and abnormal communication state features as negative samples. Learn the feature differences between normal and abnormal states through a contrastive loss function to obtain anomaly perception features. S435, Multi-feature integrated stitching: The graph convolution features, cross-modal fusion features and anomaly perception features are stitched together in the channel dimension to obtain integrated features; S436, Multilayer Perceptron Transformation: The integrated features are transformed nonlinearly through two fully connected layers, with ReLU activation function for both, to obtain the third-level features.

8. The method for multi-interface wireless communication of semiconductor equipment based on edge computing according to claim 1, characterized in that: In step S4, the wireless communication optimization model training process is as follows: S441. Dataset partitioning: The standardized training dataset is randomly divided into training set, validation set and test set according to a preset ratio, and the partitioning process maintains a consistent distribution of data in each category. S442. Transfer learning pre-training: Use publicly available industrial communication datasets to pre-train the model, set the number of pre-training rounds and the learning rate, and initialize the model parameters; S443, Optimizer and Learning Rate Settings: The AdamW optimizer is used, and the initial learning rate and weight decay coefficient are set to 0. The cosine annealing learning rate scheduling strategy is adopted. S444, Multi-task joint loss function: Training is performed using a multi-task joint loss function. The total loss is the weighted sum of the mean squared error loss of communication parameter prediction, the binary cross-entropy loss of anomaly detection, and the cross-entropy loss of protocol classification. S445, Mixed Precision Training: Employs FP16 mixed precision training technology to accelerate the model training process and reduce GPU memory usage; S446. Early stopping mechanism: When the validation set loss no longer decreases after a preset number of consecutive rounds, the early stopping mechanism is triggered to stop training and save the current optimal model parameters. S447. Model Performance Evaluation: The trained model is evaluated using a test set. Evaluation metrics include mean absolute error, root mean square error, accuracy, precision, recall, and F1 score. The model is considered qualified when all evaluation metrics meet the preset standards.

9. The method for multi-interface wireless communication of semiconductor equipment based on edge computing according to claim 1, characterized in that: In step S7, the optimal communication parameter configuration vector includes: wireless communication protocol selection, channel switching, transmit power adjustment, transmission rate adjustment, data packet size adjustment, retransmission count adjustment, modulation method adjustment, coding method adjustment, and time slot allocation adjustment.

10. A method for multi-interface wireless communication of semiconductor equipment based on edge computing according to claim 1, characterized in that: In step S9, the specific process of incremental training and model update is as follows: S91. Federated Learning Cluster Construction: Form a federated learning cluster by combining all edge computing nodes in the same workshop and designate the cloud server as the aggregation node; S92, Local Model Training: Each edge computing node uses locally collected new communication data to incrementally train the model; S93, Encrypted Gradient Upload: Each edge computing node uses homomorphic encryption technology to encrypt the local model gradient, and uploads the encrypted gradient to the cloud server once a day; S94, Global Gradient Aggregation: The cloud server uses a federated averaging algorithm to aggregate the encrypted gradients uploaded by all edge computing nodes to obtain the global model gradient; S95, Global Model Update: The cloud server updates the global model parameters based on the global model gradient and distributes the updated global model parameters to all edge computing nodes; S96. Local Model Hot Update: Edge computing nodes can update their local models using hot update technology without interrupting normal communication services, thus achieving continuous model optimization.

Images