Semiconductor wet cleaning equipment fault prediction method
Through the combination of Bayesian neural network and multi-level model, the complex data processing problem of semiconductor wet-washing equipment is solved, accurate long-term and short-term fault prediction is achieved, production losses are reduced, and key early warning and maintenance support is provided.
Patent Information
- Application Number
- CN202411084355.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-11-12
AI Technical Summary
The prior art is difficult to effectively process complex, noise, time-dependent and high-dimensional data of semiconductor wet-washing equipment, resulting in inaccurate fault prediction results and inability to meet long-term and short-term prediction needs.
Using Bayesian neural network combined with stacked sparse noise reduction automatic encoder, densely connected convolutional neural network-long and short-term memory network hybrid model and deep Q network, data processing and prediction are carried out through multi-level and multi-module combination models, including data Preprocessing, feature extraction and model optimization provide long-term and short-term fault prediction results.
Accurate and reliable fault prediction of semiconductor wet-washing equipment is achieved, losses in production process are reduced, and the accuracy and reliability of predicted results are improved, helping engineers to warn and maintain in advance in time.
Smart Images

Figure CN118940121A_ABST
Abstract
Description
[Technical Field]
[0001] This invention belongs to the field of equipment failure prediction technology, and in particular relates to a method for predicting failures in semiconductor wet cleaning equipment. [Background Technology]
[0002] Various contaminants, such as particles, organic matter, metals, and oxides, may exist on the surface of semiconductor wafers, significantly impacting device quality and yield. Therefore, wet cleaning equipment is designed to clean the wafer surface. In semiconductor manufacturing, wet cleaning is the most frequently repeated process, with multiple wet cleaning units on the production line cleaning wafers at different stages of production. If a wet cleaning unit malfunctions and requires shutdown for repair, the massive damage to wafers can result in losses amounting to tens of millions of dollars. Therefore, ensuring the safe operation of the wet cleaning process is crucial. Consequently, fault monitoring and prediction of wet cleaning equipment are critical tasks. Fault monitoring and prediction enable precise quantitative analysis of equipment health, allowing for proactive monitoring, prediction, and management of equipment health, accurately identifying the root cause of faults, generating maintenance plans based on the fault causes, and improving maintenance efficiency. However, fault prediction for wet cleaning equipment requires processing data from multiple sensors, including those for temperature, chemical concentration, flow rate, and pressure. These data exhibit complex nonlinear relationships, high data dimensionality, and complex interactions between variables, increasing the difficulty of data modeling and analysis.
[0003] The UAV fault prediction method disclosed in Chinese Patent Publication No. CN 117369403 A, based on flight data, can effectively analyze and mine combined common UAV faults and output data to determine key indicators affecting UAV stability and potential fault influencing factors. The method preprocesses, extracts, and normalizes the raw data before feeding it into a neural network for learning. A sparse pruning and denoising autoencoder is used to reduce noise and denoise the data. The Adam algorithm is then used to optimize the loss function within the network, and techniques such as Dropout, Batch Normalization, and L1 / L2 regularization are employed for testing, optimization, and lightweight adjustments to achieve a satisfactory accuracy and stability. However, this prediction method can only predict fault states within a short period based on the model's fault prediction technology. Furthermore, with a large number of predictions, error accumulation can occur, leading to inaccurate prediction results.
[0004] Therefore, a module for processing time-series data needs to be added. For example, the load point reliability assessment optimization method and device disclosed in Chinese Patent Publication No. CN117669809A uses an autoencoder to extract data features and a combination of LSTM-CNN models to capture the temporal and spatial features of power data. It uses graph neural networks combined with the power system topology for fault detection and location, and Bayesian networks to conduct in-depth analysis of risk factors, assisting decision-makers in understanding key risk points. However, this solution still has the following problems. On the one hand, due to the accuracy limitations of sensors on wet processing equipment and environmental interference, there is noise and outliers in the data. These factors may mask the true fault signals, requiring effective noise and anomaly processing. However, the current solution's processing is relatively basic and cannot meet the processing needs of complex data on wet processing equipment. On the other hand, the solution first uses LSTM to extract temporal features and then inputs the temporal features into CNN for spatial feature extraction. This algorithm model has shortcomings in high-noise data processing, high-dimensional data processing, and multimodal feature extraction, and is not suitable for fault prediction of semiconductor wet processing equipment.
[0005] Therefore, it is necessary to provide a method for predicting the failure of semiconductor wet cleaning equipment to solve the above-mentioned technical problems. [Summary of the invention]
[0006] The main objective of this invention is to provide a fault prediction method for semiconductor wet cleaning equipment, which can better cope with the complexity, noise, time dependence and high dimensionality of wet equipment data, and can also provide accurate, reliable and short-term fault prediction results, thereby reducing losses in the semiconductor wafer production process.
[0007] This invention achieves the above objective through the following technical solution: a method for predicting faults in semiconductor wet cleaning equipment, comprising the following steps:
[0008] S100: Collect data on the current operating status of the wet processing equipment;
[0009] S200: Input the data collected in step S100 into the fault prediction model to perform fault prediction;
[0010] S300. The prediction results of step S200 are evaluated using a Bayesian neural network, providing the probability distribution and confidence interval of the prediction results.
[0011] S400, output the prediction results and the probability distribution of the prediction results;
[0012] Specifically, the steps for building a fault prediction model are as follows:
[0013] S1. Collect multiple key parameter data in the wet process to form multiple corresponding sample data. Each sample data includes safe operation status data of the wet equipment and several fault status data.
[0014] S2. Perform preprocessing and standardization on the sample data. Preprocessing includes removing outliers or adding missing values. Standardization is to normalize each data point to the range of [0,1].
[0015] S3. The sample data is classified into training samples and test samples, and both the training samples and the test samples include the safe operation status data and several fault status data.
[0016] S4. Determine the data with a small number of labels in each of the training samples as minority class samples, and use synthetic minority oversampling technology to synthesize data to increase the amount of data in the minority class samples, so that each minority class sample reaches balance and thus obtains balanced training samples.
[0017] S5. A stacked sparse denoising autoencoder is used to identify the data of the balanced training samples in step S4, and the data is denoised and key features are extracted to generate a feature-based training dataset.
[0018] S6. Using a densely connected convolutional neural network-long short-term memory network hybrid model, local feature extraction is performed on the feature training dataset obtained in step S5 to generate DenseNet-LSTM hybrid features, construct a DenseNet-LSTM hybrid model, and generate a preliminary fault prediction model.
[0019] S7. A deep Q-network is used to adjust the parameters in real time according to the changes in the device status. The parameters and strategies of the DenseNet-LSTM hybrid model are optimized through a reinforcement learning mechanism, thereby optimizing the preliminary fault prediction model in step S6, obtaining the optimized fault prediction model, and using the data of the test sample to verify the fault prediction model.
[0020] S8. The prediction results of the fault prediction model are classified using the Softmax activation function.
[0021] Furthermore, the key parameter data acquisition device includes several control sensors or detection sensors distributed on the wet processing equipment.
[0022] Furthermore, the step of using the synthetic minority oversampling technique to synthesize data to increase the amount of minority class sample data is as follows:
[0023] S41. For each minority class sample, calculate its Euclidean distance to the same class sample, select the nearest multiple neighbors, randomly select one neighbor from the multiple neighbors, and randomly interpolate along the line connecting the two to generate synthetic data.
[0024] S42. Insert the synthetic data into the minority class samples to form balanced training samples.
[0025] Furthermore, the key feature extraction step of the stacked sparse noise reduction autoencoder for the data is as follows:
[0026] S51. First, construct the first layer autoencoder model. The input layer receives the balanced training samples of the training samples. Use a Gaussian noise layer to add noise to the input data, compress the high-dimensional input data to a low dimension for dimensionality reduction, and thus construct the first encoding layer. Standardize the output of the first encoding layer, reconstruct the low-dimensional data back to the original input dimension, and thus construct the first decoding layer. After standardizing the output of the first decoding layer, use the sigmoid activation function to output the reconstructed data.
[0027] S52. The safe operation status data in the training samples is used as input to train the first layer autoencoder model. The first layer autoencoder model will return the mean square error value of the corresponding sample and compare it with the set threshold to analyze the importance of the features and extract key features.
[0028] S53. Construct a second-layer autoencoder model. Use the output of the first encoding layer of the first-layer autoencoder model as the input of the second-layer autoencoder. Add noise to the input data using a Gaussian noise layer to compress the high-dimensional input data to a low dimension for dimensionality reduction, thereby constructing the second encoding layer. Standardize the output of the second encoding layer and reconstruct the low-dimensional data back to the original input dimension to construct the second decoding layer. After standardizing the output of the second decoding layer, use the sigmoid activation function to output the reconstructed data.
[0029] S54. Using the output data of the first coding layer of the first layer autoencoder model, train the second layer autoencoder model to further compress and extract key features.
[0030] S55. Repeat steps S53 and S54 in sequence to build and train more layers of autoencoder models layer by layer, using the output of the previous layer encoder model as the input of the next layer encoder model, and gradually extracting higher-level key features.
[0031] S56. After the unsupervised learning in steps S51 to S55 above, supervised learning is performed using the fault state data in the training samples to fine-tune the parameters of the multilayer autoencoder model in order to optimize the ability to extract key features.
[0032] S57. Extract feature matrices from the encoding layer of each autoencoder model, concatenate or superimpose the feature matrices extracted from each layer to form a unified high-dimensional feature matrix, reshape the extracted high-dimensional feature matrix, and generate the feature-based training dataset.
[0033] Furthermore, L1 regularization is used to implement sparsity constraints during the training of several layers of autoencoder models.
[0034] Furthermore, the method for processing data using the DenseNet-LSTM hybrid model is as follows:
[0035] S61. The densely connected convolutional neural network transmits and reuses the features of the feature-based training dataset to extract local features of the data.
[0036] S62. The Long Short-Term Memory network further extracts time-series features from the local features extracted by the densely connected convolutional neural network to generate DenseNet-LSTM hybrid features.
[0037] S63. Based on the DenseNet-LSTM hybrid features, construct the DenseNet-LSTM hybrid model, identify and classify the safe operation type and fault type of the device, and generate the preliminary fault prediction model;
[0038] S64. Based on the preliminary fault prediction model, optimize, test and lightweight adjust the preliminary fault prediction model. When compiling the model, use the Adam optimizer to optimize the sparse classification cross-entropy loss function.
[0039] Furthermore, the mean squared error loss function is used to train the deep Q network, and the deviation between the current deep Q network's predicted values and the actual values is quantified.
[0040] Compared with existing technologies, the beneficial effects of the semiconductor wet cleaning equipment fault prediction method of this invention are as follows: Through the collaborative work of SSDAE, DenseNet-LSTM, DQN, and BNN, it provides comprehensive and powerful data processing capabilities. SSDAE can effectively reduce noise and extract key features from complex and high-dimensional data on wet cleaning equipment. DenseNet-LSTM combines processing spatial and temporal series features to achieve long-term and short-term prediction and evaluation of wet cleaning equipment. DQN dynamically optimizes the model strategy to adapt to the imbalance of data, while BNN provides uncertainty estimation for the prediction results. This multi-level and multi-module combined model can better cope with the complexity, noise, time dependence, and high dimensionality of wet cleaning equipment data. It can also provide accurate, reliable, and long-term and short-term fault prediction results, helping engineers to provide early warnings and timely maintenance, ensuring the smooth operation of the production process. It has important application value and reduces losses in the semiconductor wafer production process.
Brief Description of the Drawings
[0041] Figure 1 This is a schematic diagram of the fault prediction method for semiconductor wet cleaning equipment according to an embodiment of the present invention; [Specific implementation method]
[0042] Please refer to Figure 1 This embodiment is a fault prediction method for semiconductor wet cleaning equipment. The fault prediction method for semiconductor wet cleaning equipment includes the following steps:
[0043] S100: Collect data on the current operating status of the wet processing equipment;
[0044] S200: Input the data collected in step S100 into the fault prediction model to perform fault prediction;
[0045] S300. The prediction results of step S200 are evaluated using a Bayesian neural network (BNN), providing the probability distribution and confidence interval of the prediction results.
[0046] S400, output the prediction results and the probability distribution of the prediction results.
[0047] Specifically, the steps for building a fault prediction model are as follows:
[0048] S1. Collect multiple key parameter data in the wet process to form multiple corresponding sample data. Each sample data includes safe operating status data of the wet equipment and several fault status data. The key parameter data acquisition devices include several control sensors or detection sensors distributed on the wet equipment. Key parameters include parameters related to the wet equipment such as temperature, chemical concentration, flow rate, and pressure.
[0049] S2. Preprocess and standardize the sample data. Check for missing values, delete features with excessively high missing ratios, and fill in the remaining missing values with the mean to improve data quality. Standardize the data to normalize the numerical range to [0,1] to facilitate neural network processing, reduce the computational load of the neural network, and thus speed up training. The features are transformed by scaling each feature within a given range. The estimator is scaled by the range of each feature to make it fall within the given range of the training samples, i.e., converge to the [0,1] range. The standardization formula is: Where x represents the data before standardization, and x' represents the data after standardization. min x max These are the minimum and maximum values of each training sample or each test sample, respectively.
[0050] S3. Classify the sample data to form training samples and test samples. The data of both training samples and test samples include safe operation status data and some fault status data.
[0051] S4. For training samples, the data with the fewest labels in each sample are identified as minority class samples. The Synthetic Minority Oversampling (SMOTE) technique is used to synthesize data to increase the amount of minority class samples, so that each sample data reaches balance and thus obtains balanced training samples. In fault prediction models, sample imbalance is a common problem. For example, fault state data usually only accounts for a small part of the real samples, that is, the ratio of safe operating state data to fault state data in each sample data is seriously unbalanced. In order to solve this problem, oversampling is usually used to synthesize data to increase the amount of minority class samples. In this embodiment, the Synthetic Minority Oversampling (SMOTE) technique is used to synthesize minority class samples to increase their number in the dataset, so as to achieve sample balance, making the preprocessing process more comprehensive and more adaptable.
[0052] S5. The stacked sparse denoising autoencoder (SSDAE) is used to identify the data of the balanced training samples in step S4, and the data is denoised and key features are extracted to generate a feature-based training dataset.
[0053] S6. Using a densely connected convolutional neural network (DenseNet)-long short-term memory network (LSTM) hybrid model, local feature extraction is performed on the feature training dataset obtained in step S5 to generate DenseNet-LSTM hybrid features, construct the DenseNet-LSTM hybrid model, and generate a preliminary fault prediction model.
[0054] S7. A deep Q-network (DQN) is used to adjust the parameters in real time according to the changes in the device status. The parameters and strategies of the DenseNet-LSTM hybrid model are optimized through the reinforcement learning mechanism, thereby optimizing the preliminary fault prediction model in step S6, obtaining the optimized fault prediction model, and using the test sample data to verify the fault prediction model.
[0055] S8. Use the Softmax activation function to classify and output the prediction results of the fault prediction model.
[0056] Specifically, the steps of using Synthetic Minority Oversampling (SMOTE) technique to increase the amount of minority class sample data are as follows:
[0057] S41. For each minority class sample, calculate its Euclidean distance to the same class sample, select the nearest multiple neighbors, randomly select one neighbor from the multiple neighbors, and randomly interpolate along the line connecting the two to generate synthetic data.
[0058] S42. Insert the synthetic data into the minority class samples to form balanced training samples.
[0059] Specifically, the feature extraction steps of the stacked sparse denoising autoencoder (SSDAE) are as follows:
[0060] S51. First, construct the first layer autoencoder model. The input layer receives balanced training samples. A Gaussian noise layer is used to add noise to the input data, compressing the high-dimensional input data to a low dimension for dimensionality reduction, thereby constructing the first encoding layer. The output of the first encoding layer is standardized, and the low-dimensional data is reconstructed back to the original input dimension to construct the first decoding layer. After standardizing the output of the first decoding layer, the sigmoid activation function is used to output the reconstructed data.
[0061] S52. Use the safe operating status data within the training samples as input to train the first layer autoencoder model. The first layer autoencoder model will return the mean square error value of the corresponding sample and compare it with the set threshold. The training of the first layer autoencoder model analyzes the importance of features and extracts key features.
[0062] S53. Construct a second-layer autoencoder model. Use the output of the first encoding layer of the first-layer autoencoder model as the input of the second-layer autoencoder. Add noise to the input data using a Gaussian noise layer to compress the high-dimensional input data to a low dimension for dimensionality reduction, thereby constructing the second encoding layer. Standardize the output of the second encoding layer and reconstruct the low-dimensional data back to the original input dimension to construct the second decoding layer. After standardizing the output of the second decoding layer, use the sigmoid activation function to output the reconstructed data.
[0063] S54. Using the output data of the first encoding layer of the first autoencoder model, train the second autoencoder model to further compress and extract key features.
[0064] S55. Repeat steps S53 and S54 in sequence to build and train more layers of autoencoder models layer by layer, using the output of the previous layer encoder model as the input of the next layer encoder model, and gradually extracting higher-level key features.
[0065] S56. After the unsupervised learning in steps S51 to S55 above, supervised learning is performed using fault state data within the training samples to fine-tune the parameters of the multilayer autoencoder model in order to optimize the ability to extract key features.
[0066] S57. Extract feature matrices from the encoding layers of each autoencoder model, concatenate or superimpose the feature matrices extracted from each layer to form a unified high-dimensional feature matrix, reshape the extracted high-dimensional feature matrix, and generate a feature-based training dataset.
[0067] Stacked Sparse Denoising Autoencoders (SSDAEs) consist of multiple denoising autoencoder models stacked together. Each layer progressively learns higher-level feature representations of the data, and the output of one layer can be used as input to train the next layer. The entire model forms a deep neural network, which helps to capture complex structures and features in the data. Moreover, by introducing noisy data during training, SSDAEs can learn how to extract stable features in the presence of noise, enabling accurate feature extraction from noisy data in wet processing equipment and improving the robustness and noise resistance of the model.
[0068] Sparsity constraints are incorporated into the training of several layers of the autoencoder model to encourage the learning of sparser data representations. This helps to highlight important features, improves generalization ability, and makes the extracted features more distinctive and representative. A sparsity regularization term is added to the loss function of the autoencoder. In this embodiment, L1 regularization is used. During training, the addition of the L1 regularization term makes the dummy weights zero, thereby achieving feature selection and sparsity. This means that only a few neurons play an important role in encoding the input data, while the weights of the remaining neurons are compressed to near-zero values, thus achieving high compression and sparse representation of data features.
[0069] Therefore, the Stacked Sparse Denoising Auto Encoder (SSDAE) uses a deep learning structure to denoise and extract features from data, which can capture complex nonlinear features in high-dimensional data. At the same time, the sparseness constraint helps to extract key features, and it can effectively capture data with complex nonlinear relationships, high data dimensionality, and complex interactions between variables in wet processing equipment.
[0070] Specifically, the method for processing data using the DenseNet-LSTM hybrid model is as follows:
[0071] S61. Densely connected convolutional neural networks (DenseNet) are used to transfer and reuse features in the feature training dataset, and extract local features of the data. Densely connected convolutional neural networks (DenseNet) can transfer and reuse features between multiple layers, which enhances the model's ability to capture complex relationships between multiple variables. It can extract local features from data with complex nonlinear relationships, high data dimensionality, and complex interactions between variables in wet processing equipment.
[0072] S62. A Long Short-Term Memory (LSTM) network is used to further extract time-series features from the local features extracted by the densely connected convolutional neural network, generating DenseNet-LSTM hybrid features. Since the data has strong time-series characteristics, there may be significant dependencies between different time points. The Long Short-Term Memory (LSTM) network is good at processing time-series data and can effectively capture the long-short-term dependencies in the data. Therefore, the Long Short-Term Memory (LSTM) network is used to capture and process these time-series data.
[0073] S63. Based on the DenseNet-LSTM hybrid features, a DenseNet-LSTM hybrid model is constructed to identify and classify the safe operation type and fault type of the equipment, and generate a preliminary fault prediction model. By combining the spatial feature extraction of the densely connected convolutional neural network (DenseNet) and the temporal feature capture capability of the long short-term memory network (LSTM), the preliminary fault prediction model can comprehensively handle the complex patterns in time series data. Therefore, it can achieve long-term and short-term fault prediction for wet processing equipment.
[0074] S64. Based on the preliminary fault prediction model, the model is optimized, tested, and lightweighted. The Adam optimizer is used during model compilation to optimize the sparse classification cross-entropy loss function.
[0075] During optimization, the fault prediction model is tested and validated to ensure that the training process has clear objectives and evaluation criteria, thereby helping to optimize the model's performance. The Adam optimizer is used during model compilation to optimize the sparse classification cross-entropy loss function.
[0076] Specifically, the steps for constructing the DenseNet-LSTM hybrid model are as follows:
[0077] S631, The input layer of the densely connected convolutional neural network (DenseNet) receives the feature matrix reshaped by the stacked sparse denoising autoencoder (SSDAE);
[0078] S632. Densely connected convolutional neural networks (DenseNet) extract features from the feature matrix using dense blocks. Each dense block consists of multiple convolutional layers, and the output of each layer in the dense block serves as the input to the next layer, progressively extracting higher-level feature representations, which can improve feature extraction efficiency and accuracy. Moreover, the ReLU activation function is used after each convolutional layer, which can effectively introduce nonlinearity, avoid the gradient vanishing problem, improve computational efficiency, and enhance the generalization ability of the model through sparse activation.
[0079] S633, The transition layer of the densely connected convolutional neural network (DenseNet) reduces the dimension of the feature map and uses maximum pooling for sampling;
[0080] S634. Use a two-layer Long Short-Term Memory (LSTM) network to capture time series features and add a dropout layer to prevent overfitting.
[0081] S635. Use a fully connected layer for feature fusion and output using the ReLU activation function;
[0082] S636. A Softmax output layer is connected after the fully connected layer to map the features output by the network to the probability of the image belonging to different categories, and the probability of different categories is classified and output.
[0083] The Deep Q-Network (DQN) module approximates the Q-value function using a neural Q-network, thereby optimizing the decision-making strategy. Since the DQN module calculates the difference between the estimated Q-value and the target Q-value under the current policy in each iteration, the Mean Squared Error (MSE) loss function is used to measure the difference between the predicted Q-value and the target Q-value y. The MSE loss function guides the update of the Q-network parameters, adjusting them by minimizing the MSE to make the predicted Q-value as close as possible to the true target Q-value. The DQN is trained using the MSE loss function, quantifying the deviation between the current prediction and the actual value. By continuously adjusting the network weights, the deviation between prediction and reality is reduced, resulting in better behavioral decisions.
[0084] During the testing and validation phase, the trained model is tested using a test sample set, and performance metrics such as accuracy, recall, and F1 score are calculated. If the test results are not ideal, the model parameters are fine-tuned and retrained to further improve the speed and accuracy of recognition.
[0085] Bayesian neural networks (BNNs) introduce uncertainty assessment into the model parameters. By outputting the probability distribution of the prediction results, they can identify abnormal data with high uncertainty, helping decision-makers understand the confidence level of the prediction and reduce the risk of misjudgment.
[0086] This invention provides comprehensive and powerful data processing capabilities through the collaborative work of SSDAE, DenseNet-LSTM, DQN, and BNN. SSDAE effectively reduces noise and extracts key features from data from wet processing equipment; DenseNet-LSTM combines spatial and temporal features to achieve short- and long-term prediction and evaluation of wet processing equipment; DQN dynamically optimizes the model strategy to adapt to data imbalance; and BNN provides uncertainty estimation for the prediction results. This multi-level, multi-module combined model can better handle the complexity, noise, time dependence, and high dimensionality of wet processing equipment data, providing accurate and reliable fault prediction results. It helps engineers provide early warnings and timely maintenance, ensuring the smooth operation of the production process, and has significant application value and promising prospects for widespread adoption.
[0087] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.
Claims
1. A method for predicting faults in semiconductor wet cleaning equipment, characterized in that: It includes the following steps: S100, collecting data of the current operating state of the wet process equipment; S200, inputting the data collected in step S100 into a fault prediction model to perform fault prediction; S300, using a Bayesian neural network to evaluate the prediction result of step S200, and provide a probability distribution and a confidence interval of the prediction result; S400, outputting the prediction result and the probability distribution of the prediction result; The steps of building a fault prediction model are: S1. Collect multiple key parameter data in the wet process to form corresponding multiple sample data, each of which includes safe operation status data of the wet process equipment and several fault status data; S2. Preprocessing and standardizing the sample data. Preprocessing includes removing outliers or supplementing missing values. Standardization is to normalize each data to the range of [0, 1]. S3, classifying the sample data into training samples and test samples, wherein the training samples and the test samples both include the safe operation status data and a number of the fault status data; S4, determining the data with a small number of labels in each sample data of the training samples as minority samples, using synthetic minority oversampling technology to synthesize data to increase the data volume of the minority samples, so that each of the minority samples reaches a balance to obtain a balanced training sample; S5, using a stacked sparse denoising autoencoder to identify the data of the balanced training samples in step S4, and perform denoising and key feature extraction on the data to generate a feature-based training data set; S6, using a densely connected convolutional neural network-long short-term memory network hybrid model, extracting local features of the characterized training data set obtained in step S5, generating DenseNet-LSTM hybrid features, constructing a DenseNet-LSTM hybrid model, and generating a preliminary fault prediction model; S7, using a deep Q network to adjust parameters in real time according to changes in device status, and optimizing the parameters and strategies of the DenseNet-LSTM hybrid model through a reinforcement learning mechanism, thereby optimizing the preliminary fault prediction model of step S6, obtaining an optimized fault prediction model, and using the data of the test sample to verify the fault prediction model; S8. Use the Softmax activation function to classify the prediction results of the fault prediction model.
2. The semiconductor wet cleaning equipment failure prediction method according to claim 1, characterized in that: The key parameter data acquisition device includes a plurality of control sensors or detection sensors distributed on the wet process equipment.
3. The semiconductor wet cleaning equipment failure prediction method according to claim 1, characterized in that: The steps of synthesizing data using the synthetic minority oversampling technique to increase the amount of minority class sample data are as follows: S41, for each of the minority class samples, calculate the Euclidean distance between the sample and the same class sample, select multiple nearest neighbors, randomly select a neighbor from the multiple neighbors, and randomly interpolate along the line connecting the two to generate synthetic data; S42: inserting the synthetic data into the minority class samples to form balanced training samples.
4. The semiconductor wet cleaning equipment failure prediction method according to claim 3, characterized in that: The steps of extracting key features of data by the stacked sparse denoising autoencoder are as follows: S51, first construct a first-layer autoencoder model, where the input layer receives the balanced training sample of the training sample, uses a Gaussian noise layer to add noise to the input data, compresses the input high-dimensional data to a low dimension for dimensionality reduction processing to construct a first encoding layer, standardizes the output of the first encoding layer, reconstructs the low-dimensional data back to the original input dimension to construct a first decoding layer, and standardizes the output of the first decoding layer and outputs the reconstructed data using a sigmoid activation function; S52, using the safe operation status data in the training sample as input, training the first-layer autoencoder model, the first-layer autoencoder model returns the mean square error value of the corresponding sample and compares it with the set threshold, analyzes the importance of the features and extracts key features; S53, constructing a second-layer autoencoder model, using the first encoding layer output of the first-layer autoencoder model as the input of the second-layer autoencoder, using a Gaussian noise layer to add noise to the input data, compressing the input high-dimensional data to a low dimension for dimensionality reduction processing to construct a second encoding layer, standardizing the output of the second encoding layer, reconstructing the low-dimensional data back to the original input dimension to construct a second decoding layer, and standardizing the output of the second decoding layer and outputting the reconstructed data using a sigmoid activation function; S54, using the first encoding layer output data of the first layer autoencoder model to train the second layer autoencoder model to further compress and extract key features; S55, repeating the above steps S53 and S54 in sequence, constructing and training more layers of autoencoder models layer by layer, using the output of the encoder model of the previous layer as the input of the encoder model of the next layer, and gradually extracting higher-level key features; S56, after the unsupervised learning in the above steps S51 to S55, supervised learning is performed using the fault state data in the training sample to fine-tune the parameters of the multi-layer autoencoder model to optimize the ability to extract key features; S57, extracting a feature matrix from the encoding layer of each layer of the autoencoder model, concatenating or superimposing the feature matrices extracted from each layer to form a unified high-dimensional feature matrix, reshaping the extracted high-dimensional feature matrix, and generating the characterized training data set.
5. The semiconductor wet cleaning equipment failure prediction method according to claim 4, characterized in that: L1 regularization is used to implement sparsity constraints when training several layers of autoencoder models.
6. The semiconductor wet cleaning equipment failure prediction method according to claim 1, characterized in that: The method for processing data by the DenseNet-LSTM hybrid model is: S61, the densely connected convolutional neural network transfers and reuses the features of the characterized training data set to extract local features of the data; S62, the long short-term memory network further extracts time series features from the local features extracted by the densely connected convolutional neural network to generate DenseNet-LSTM hybrid features; S63. Based on the DenseNet-LSTM hybrid features, construct the DenseNet-LSTM hybrid model, identify and classify the safe operation type and fault type of the equipment, and generate the preliminary fault prediction model; S64. Based on the preliminary fault prediction model, the preliminary fault prediction model is optimized, tested and lightweight adjusted. When the model is compiled, the Adam optimizer is used to optimize the sparse classification cross entropy loss function.
7. The semiconductor wet cleaning equipment failure prediction method according to claim 1, characterized in that: The deep Q network is trained using a mean square error loss function to quantify the deviation between the predicted value and the actual value of the deep Q network.
Citation Information
Patent Citations
Unmanned aerial vehicle fault prediction method based on flight data
CN117369403A
Load point reliability evaluation optimization method and device
CN117669809A
