Preparation data analysis method and system for dry film photoresist

Abstract

The invention provides a dry film photoresist preparation data analysis method and system, and the method comprises the steps: obtaining the time sequence process data of the whole process of dry film photoresist preparation, carrying out the preprocessing of the time sequence process data, inputting the time sequence process data into a TCN-NP model, outputting the quality index prediction result of the dry film photoresist, constructing a stage type small sample network, and carrying out the prediction of the quality index of the dry film photoresist. Performing fault identification on the time sequence process data through a stage type small sample network to obtain a fault diagnosis result, and generating a process parameter optimization instruction according to a quality index prediction result and the fault diagnosis result in combination with a hierarchical reinforcement learning framework; through the full-flow integrated scheme including time sequence data processing, quality prediction and uncertainty quantification, small sample fault identification and collaborative optimization decision making, the problems of low process control precision and blind optimization decision making in the dry film photoresist preparation process are solved, so that the stability of the photoresist preparation quality is improved.

Description

Technical Field

[0001] This invention relates to the field of photoresist technology, and more specifically, to a method and system for analyzing data related to the preparation of dry film photoresist. Background Technology

[0002] With the rapid development of high-end fields such as microelectronics manufacturing and precision circuit processing, the market demand for dry film photoresist, as a basic material, has increased significantly, and the requirements for high precision and high stability of product quality have become increasingly stringent. The dry film photoresist preparation process is characterized by multi-stage coupling, strong temporal correlation of process parameters, and complex quality-influencing factors; therefore, the quality of process data analysis directly determines product performance. To achieve control over the dry film photoresist preparation process, quality prediction, fault identification, and process optimization through full-process process data analysis have become a key research focus in the field of photoresist production operation and quality control. However, existing related technologies often have many shortcomings. Existing dry film photoresist fabrication data analysis techniques mostly rely on single-dimensional data or local process parameters as the basis for analysis, failing to fully consider the correlation between all process steps and the coupling effect of multi-dimensional time-series data. On the other hand, existing analysis models are mostly general-purpose data processing or prediction algorithms, such as single LSTM or ARIMA, which cannot accurately capture long-term process dependencies and are difficult to quantify prediction uncertainties. At the same time, in the fault identification and process optimization stages, either the reliance on a large number of fault samples makes the identification scheme difficult to implement, or the optimization decision ignores the synergy between quality prediction and fault diagnosis, leading to process fluctuations.

[0003] Therefore, there is an urgent need for an integrated technical solution that can achieve accurate processing of time-series data throughout the entire process, quality prediction and uncertainty quantification, small-sample fault identification, and collaborative optimization decision-making, in order to solve the problems of low process control precision and blind optimization decision-making in the dry film photoresist preparation process, thereby improving the stability of photoresist preparation quality. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for analyzing data from the preparation of dry film photoresist, the method comprising: Acquire the timing process data of the entire dry film photoresist preparation process, and preprocess the timing process data; Input the timing process data into the TCN-NP model to output the quality index prediction results of the dry film photoresist; A staged few-sample network is constructed, and fault identification is performed on time-series process data based on the staged few-sample network to obtain fault diagnosis results; Based on the predicted results of quality indicators and the results of fault diagnosis, and combined with a hierarchical reinforcement learning framework, process parameter optimization instructions are generated.

[0005] As a further aspect of the present invention, timing process data of the entire dry film photoresist preparation process is obtained, and the timing process data is preprocessed, including: Based on distributed acquisition, raw material supply parameters, stirring parameters, coating speed, drying temperature, curing time, and ambient temperature and humidity are obtained and combined to form the time-series process data of the entire dry film photoresist preparation process. The time-series process data is preprocessed by cleaning and alignment to obtain standardized time-series process data. The cleaning includes removing outliers and filling in missing values ​​in the time-series process data. The alignment refers to unifying the time scale of the time-series process data in all dimensions.

[0006] As a further aspect of the present invention, timing process data is input into the TCN-NP model, and the predicted quality indicators of the dry film photoresist are output, including: A TCN-NP model is constructed based on temporal convolutional networks and neural processes. The TCN-NP model includes a feature extraction layer and a probabilistic inference layer. Among them, temporal convolutional networks are used as deterministic feature extractors to replace multilayer perceptrons in neural processes; Time-series process data is input into the feature extraction layer, which extracts high-dimensional time-series feature vectors based on a dilated causal convolution structure. The high-dimensional temporal feature vector is input into the probabilistic inference layer. Probability calculation is performed based on the latent variable framework of the probabilistic inference layer, and the predicted value of the dry film photoresist quality index and the 95% dynamic confidence interval are output. Assisted decision-making is based on a 95% dynamic confidence interval to obtain decision support information; The quality index prediction results are composed of the predicted values ​​of dry film photoresist quality indicators, the 95% dynamic confidence interval, and decision support information.

[0007] As a further aspect of the present invention, decision support information is obtained based on a 95% dynamic confidence interval, including: Set a width threshold and monitor the width of the 95% dynamic confidence interval in real time; If the width of the obtained 95% dynamic confidence interval is greater than the width threshold, it is determined that the current process state exceeds the empirical range, and the corresponding decision support information is output simultaneously. If the width of the obtained 95% dynamic confidence interval is less than or equal to the width threshold, the current process state is determined to be within the empirical range, and blank decision support information is output.

[0008] As a further aspect of the present invention, a staged few-sample network is constructed, and fault identification is performed on time-series process data based on the staged few-sample network to obtain fault diagnosis results, including: Construct a staged few-shot network comprising a first stage and a second stage, and train the staged few-shot network. The first stage maps the time-series process data into real-time feature vectors based on the feature encoder, and the second stage performs fault type calibration on the time-series process data based on the feature encoder. Set a distance threshold. If the cosine distance between the real-time feature vector and the normal feature vector is less than or equal to the distance threshold, it is determined to be normal process data. If the cosine distance between the real-time feature vector and the normal feature vector is greater than the distance threshold, it enters the second stage. Obtain the cosine distance between the real-time feature vector and the vectors of each fault type in the fault feature library, and filter out the minimum cosine distance; Set a fault matching threshold. If the minimum cosine distance is less than or equal to the fault matching threshold, then the fault type corresponding to the minimum cosine distance is taken as the fault diagnosis result. If the minimum cosine distance is greater than the fault matching threshold, the current fault is determined to be a completely new fault type, and an incremental learning mechanism is triggered.

[0009] As a further aspect of the present invention, training a staged few-shot network includes: Obtain normal process data and train the feature encoder based on the normal process data; The feature encoder is frozen, and fault process data with fault labels is obtained. The fault process data is input into the frozen feature encoder to generate fault sample feature vectors. The fault sample feature vectors are calibrated to obtain fault type vectors. A fault feature library is constructed based on the fault type vectors.

[0010] As a further aspect of the present invention, based on the quality index prediction results and fault diagnosis results, and combined with a hierarchical reinforcement learning framework, process parameter optimization instructions are generated, including: Define the process regulation subtask, introduce the SAC algorithm to the high-level policy controller and the low-level policy controller, and build a hierarchical reinforcement learning framework based on the high-level policy controller and the low-level policy controller. Based on the process adjustment subtask, the quality indicator prediction results and fault diagnosis results are input into the high-level strategy controller, and the process option instructions are output. The process option instructions are input into the underlying strategy controller, and process parameter optimization instructions are output.

[0011] Furthermore, embodiments of the present invention also provide a data analysis system for the preparation of dry film photoresist, comprising: The acquisition module is used to acquire the timing process data of the entire dry film photoresist preparation process; The processing module is used to preprocess the time-series process data, generate process parameter optimization instructions based on the quality index prediction results and fault diagnosis results, and in combination with the hierarchical reinforcement learning framework. The input module inputs the timing process data into the TCN-NP model; The output module outputs the quality index prediction results of the dry film photoresist based on the TCN-NP model; The building module is used to build a staged few-shot network; The identification module identifies faults in time-series process data based on a staged few-sample network and obtains fault diagnosis results. Attached Figure Description

[0012] Figure 1 This is a flowchart of the steps in the data analysis method for the preparation of dry film photoresist according to the present invention; Figure 2 This is a flowchart of the steps for obtaining fault diagnosis results in a data analysis method for the preparation of dry film photoresist according to the present invention; Figure 3 This is a schematic diagram of a dry film photoresist preparation data analysis system according to the present invention. Detailed Implementation

[0013] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0014] like Figure 1 As shown, a method for analyzing data from the preparation of dry film photoresist includes the following steps: Step S1: Obtain the timing process data of the entire dry film photoresist preparation process, and preprocess the timing process data.

[0015] Step S1 includes the following: Based on distributed acquisition, raw material supply parameters, stirring parameters, coating speed, drying temperature, curing time, and ambient temperature and humidity are obtained and combined to form the time-series process data of the entire dry film photoresist preparation process.

[0016] Specifically, independent data acquisition nodes can be set up in process stages such as raw material supply, stirring, coating, drying, curing, and environmental control. Raw material supply parameters include raw material supply rate, supply pressure, and raw material viscosity, which directly affect the stability of the basic components of the photoresist. Stirring parameters include stirring speed, stirring time, and stirring power, which determine the uniformity of raw material mixing. Coating speed, drying temperature, and curing time are parameters that affect the quality of dry film formation. Ambient temperature and humidity indirectly affect process stability and product performance. Each data acquisition node transmits the collected data to a central data server in real time through an industrial communication network, such as Modbus TCP, and integrates it according to timestamps to form time-series process data that includes the entire process, ensuring that the data can completely reflect the entire process status of dry film photoresist preparation.

[0017] In some possible embodiments, taking the film thickness uniformity optimization batch of a dry film photoresist manufacturer as an example, a distributed acquisition method is used to deploy 6 acquisition nodes. This includes deploying pressure and rate sensors in the raw material supply stage to collect raw material supply pressure and supply rate; deploying speed and power sensors in the stirring stage to collect stirring speed and stirring power; deploying speed sensors in the coating stage to collect coating speed; deploying temperature sensors in the drying stage to collect drying temperature; deploying time and temperature sensors in the curing stage to collect curing time and curing temperature; and deploying temperature and humidity sensors in the environmental area to collect ambient temperature and humidity. The data collected by each node is transmitted to the central server via industrial Ethernet and correlated according to the same timestamp to form full-process time-series process data. For example, the data at 9:00:00 includes: raw material supply pressure 0.4MPa, supply rate 5kg / h, stirring speed 530r / min, coating speed 30mm / s, drying temperature 87℃, curing temperature 120℃, and ambient humidity 48%RH, etc.

[0018] The time-series process data is preprocessed by cleaning and alignment to obtain standardized time-series process data. The cleaning includes removing outliers and filling in missing values ​​in the time-series process data. The alignment refers to unifying the time scale of the time-series process data in all dimensions.

[0019] It should be noted that in the photoresist production process, sensor malfunctions, electromagnetic interference, and human error often occur, causing data to deviate from the normal process range and generate outliers. In such cases, the collected data needs to be cleaned, using methods such as the 3σ criterion, quartile method, or process rule method, which can accurately identify and remove outliers. The photoresist production process also involves issues such as temporary sensor offline time and data transmission delays, resulting in missing values. Methods such as linear interpolation and mean imputation can be used to fill in missing values ​​while ensuring the temporal correlation of the data, avoiding the impact of incomplete data on subsequent analysis. Furthermore, different sensors have inconsistent sampling frequencies. A unified time scale, i.e., a method to determine the target sampling frequency, is needed. High-sampling-frequency data can be downsampled using methods such as averaging or maximizing, while low-sampling-frequency data can be upsampled using methods such as linear interpolation. This ensures that the temporal process data of each dimension correspond at the same time point, thereby obtaining a standardized temporal process data matrix.

[0020] In some possible embodiments, continuing the above example, the collected full-process time-series process data is preprocessed. This includes analyzing the coating speed data using the 3σ criterion to find a value of 38 mm / s. Given that the normal coating speed range is 25-35 mm / s, this value exceeds the mean ± 3σ range and is determined to be caused by sensor electromagnetic interference. Therefore, the 38 mm / s coating speed data is identified as an outlier and deleted. The raw material supply pressure data has missing values ​​in the time period from 10:00:00 to 10:00:20. Linear interpolation is used to find missing values ​​by interpolating values ​​before and after the time interval, such as 9:59. The interpolation values ​​for 0.39 MPa at 50 and 0.41 MPa at 10:00:30 were used to obtain a supply pressure value of 0.40 MPa for the missing time period. The alignment process included determining the target sampling frequency as 10 seconds per data point, downsampling data such as raw material supply pressure and stirring speed at a sampling frequency of 1 Hz, and upsampling data such as drying temperature and curing temperature at a sampling frequency of 0.05 Hz to form a standardized time-series process data matrix. The time-series process data matrix includes 1440 time points and 6 dimensions, with each cell corresponding to a standardized value of a parameter at a certain moment.

[0021] Step S2: Input the timing process data into the TCN-NP model and output the quality index prediction results of the dry film photoresist.

[0022] Step S2 includes the following: The TCN-NP model is constructed based on temporal convolutional networks and neural processes. The TCN-NP model includes a feature extraction layer and a probabilistic inference layer. The temporal convolutional network is used as a deterministic feature extractor to replace the multilayer perceptron in the neural process.

[0023] It should be noted that traditional neural processes often use multilayer perceptrons as feature extractors. However, multilayer perceptrons have a weak ability to acquire long-range dependencies of temporal data. In contrast, the temporal correlation of dry film photoresist preparation process parameters is strong, such as the influence of stirring parameters on the subsequent coating quality. Therefore, this step constructs a TCN-NP model including feature extraction layers and probabilistic inference layers through temporal convolutional networks and neural processes, and uses temporal convolutional networks as deterministic feature extractors to replace multilayer perceptrons in neural processes.

[0024] Specifically, the temporal convolutional network includes a dilated causal convolutional structure, which can exponentially increase the receptive field by stacking convolutional layers, thereby obtaining long-range dependencies in temporal process data; the probabilistic inference layer, through a latent variable framework with probabilistic computation capabilities in neural processes, can transform the deterministic features extracted by the temporal convolutional network into an output containing predicted values ​​and uncertainties; this step works in concert through the two-layer architecture of the feature extraction layer and the probabilistic inference layer, which not only ensures the effectiveness of feature extraction, but also quantifies the prediction uncertainty, thereby improving the reliability of quality prediction.

[0025] Understandably, photoresist process parameters change continuously over time and exhibit long-range correlations, requiring models to accurately capture temporal features. Furthermore, industrial production necessitates a clear understanding of the reliability of prediction results to support decision-making and avoid process adjustment errors due to prediction uncertainty. Traditional models struggle to meet the high requirements of photoresist manufacturing. Therefore, in photoresist fabrication processes characterized by complexity, temporality, and the need for reliable predictions in industrial settings, the TCN-NP model is employed to meet these high requirements. Compared to traditional temporal prediction models such as LSTM and GRU, it achieves more efficient long-range dependency capture through dilated causal convolution and possesses stronger parallel computing capabilities, making it suitable for real-time prediction of large-scale process data. Compared to single neural process models, using a temporal convolutional network instead of a multilayer perceptron as a feature extractor improves the accuracy of feature extraction from temporal data. Compared to traditional point prediction models, it can output dynamic confidence intervals to quantify prediction uncertainty, reducing industrial production risks.

[0026] In some possible embodiments, continuing the examples described above, a TCN-NP model is constructed, comprising a feature extraction layer and a probabilistic inference layer. The feature extraction layer employs a 4-layer temporal convolutional network structure, with each layer having a kernel size of 3 and dilation rates of 1, 2, 4, and 8, respectively. This dilated causal convolution captures the long-range dependencies between parameters such as raw material supply pressure fluctuations, stirring speed changes, and coating speed adjustments. The probabilistic inference layer uses a latent variable framework of neural processes, where the latent variables follow a Gaussian distribution, and probabilities are obtained through variational inference. During model training, standardized temporal process data from 30 historical qualified production batches of the enterprise and corresponding measured film thickness uniformity data can be used as the training set. Parameter optimization is performed by minimizing the mean square error between predicted and measured values ​​and the KL divergence of the probability distribution, thereby completing the training of the TCN-NP model.

[0027] Furthermore, the dataset is first divided. Standardized time-series process data from 30 qualified production batches in the company's history and corresponding measured film thickness uniformity data are selected as the total dataset, which can be divided into training, validation, and test sets in a 7:2:1 ratio. Then, the training parameters are configured. The Adam optimizer can be used, with the initial learning rate set to 0.001 and a learning rate decay strategy adopted, i.e., decaying to 0.9 times the original value every 20 rounds. The batch size is set to 32, the number of training rounds is set to 100, and the early stopping strategy threshold is set to 5, i.e., training stops if the loss on the validation set does not decrease for 5 consecutive rounds. The loss function is a weighted sum of mean squared error and KL divergence, with weights of 0.6 and 0.4, respectively. The mean squared error is used to optimize the deviation between the predicted and measured values, and the KL divergence is used to optimize the consistency between the latent variable distribution and the prior distribution. Finally, the trained TCN-NP model is obtained through parameter optimization.

[0028] The time-series process data is input into the feature extraction layer, which extracts high-dimensional time-series feature vectors based on the dilated causal convolution structure.

[0029] Specifically, the feature extraction layer extracts features through a dilated causal convolution structure. Causal convolution ensures that the model only uses current and past process data for feature extraction, thus conforming to the temporal logic of dry film photoresist preparation. Dilated convolution expands the receptive field of the model without increasing the number of parameters, enabling it to capture the variation patterns of process parameters over a longer time span. The feature extraction layer transforms high-dimensional, multivariate standardized temporal process data into low-dimensional, structured high-dimensional temporal feature vectors, thereby condensing process state information.

[0030] Furthermore, in the feature extraction process, the standardized time-series process data first undergoes preliminary feature extraction through a one-dimensional convolutional layer, then feature optimization is performed through a batch normalization layer and activation function, and finally iterative processing through multiple dilated causal convolutional layers to output a high-dimensional time-series feature vector.

[0031] In some possible embodiments, continuing the above example, the standardized time-series process data matrix is ​​input into the TCN feature extraction layer. First, a one-dimensional convolutional layer is used to perform preliminary feature extraction on the process parameters in six dimensions. Then, a batch normalization layer is used to eliminate data distribution differences, and nonlinear features are introduced through the ReLU activation function. Then, through four layers of dilated causal convolutional layers, the receptive field is gradually expanded to obtain long-range dependencies such as raw material supply pressure fluctuations, fine-tuning of stirring speed, and changes in coating speed. Finally, a high-dimensional time-series feature vector of 128 dimensions is output.

[0032] The high-dimensional temporal feature vector is input into the probabilistic inference layer. Probability calculation is performed based on the latent variable framework of the probabilistic inference layer, and the predicted value of the dry film photoresist quality index and the 95% dynamic confidence interval are output.

[0033] Specifically, high-dimensional time-series feature vectors are input into the probabilistic inference layer, which performs probabilistic modeling of the feature vectors using a latent variable framework.

[0034] It should be noted that the latent variable framework assumes the existence of a latent variable that follows a specific probability distribution, such as a Gaussian distribution, which can characterize the potential correlation between process data and quality indicators. Using variational inference, the model learns the posterior distribution of the latent variable and performs probability calculations based on this posterior distribution to obtain the predicted distribution of the quality indicators. Finally, based on the predicted distribution, the model outputs the predicted values ​​of the quality indicators and a 95% dynamic confidence interval. The 95% dynamic confidence interval represents a 95% probability that the actual quality indicator value falls within this interval, and the width of the 95% dynamic confidence interval is simultaneously obtained. The width of the 95% dynamic confidence interval is dynamically adjusted according to changes in the process data, and this width reflects the reliability of the prediction results.

[0035] In some possible embodiments, the 128-dimensional high-dimensional temporal feature vector obtained in step S2-2 is input into the probabilistic inference layer. The latent variables of the probabilistic inference layer follow a Gaussian distribution. The posterior distribution parameters of the latent variables are obtained through variational inference learning, with a mean of 0.2 and a variance of 0.01. Probability calculation is performed through the posterior distribution to obtain the predicted distribution of film thickness uniformity, and the predicted value of film thickness uniformity is output as ±2.6μm. Then, the 95% dynamic confidence interval is [±2.1μm, ±3.1μm]. The confidence interval indicates that there is a 95% probability that the actual film thickness uniformity of this batch of dry film photoresist falls between ±2.1μm and ±3.1μm, with an interval width of 0.5μm.

[0036] Assisted decision-making is based on a 95% dynamic confidence interval to obtain decision support information.

[0037] The content of the decision support includes: Set a width threshold and monitor the width of the 95% dynamic confidence interval in real time.

[0038] Specifically, the width threshold needs to be determined based on the company's historical production data, process requirements, and model performance. First, the 95% dynamic confidence interval widths and corresponding prediction deviation rates of multiple historical qualified and unqualified batches are statistically analyzed. With a prediction deviation rate of less than or equal to 5%, the maximum confidence interval width that meets this target is determined as the width threshold. If the company has higher requirements for prediction accuracy, the threshold can be appropriately lowered; if it needs to improve the model's fault tolerance, the threshold can be appropriately increased. The 95% dynamic confidence interval width output by the probabilistic inference layer is collected and analyzed in real time to ensure that changes in process status can be captured in a timely manner, providing real-time support for subsequent decision-making.

[0039] In some possible embodiments, statistical analysis of the 95% dynamic confidence interval width and prediction deviation rate data of 50 historical production batches revealed that when the confidence interval width is less than or equal to 0.6 μm, the prediction deviation rate of film thickness uniformity for all batches is less than or equal to 5%, meeting the company's accuracy requirement of a prediction deviation rate of less than or equal to 5%. When the width is greater than 0.6 μm, the prediction deviation rate is greater than 5%, failing to meet the accuracy requirement. Therefore, the width threshold can be set to 0.6 μm. During the production process of this batch, the 95% dynamic confidence interval width output by the probabilistic inference layer is collected every 10 seconds by the data monitoring module to achieve real-time monitoring of the width.

[0040] If the width of the obtained 95% dynamic confidence interval is greater than the width threshold, it is determined that the current process state exceeds the empirical range, and the corresponding decision support information is output simultaneously.

[0041] Specifically, when the width of the 95% dynamic confidence interval is greater than the preset threshold, it indicates that the current process data contains abnormal features that have not been encountered during the model training process, such as sudden changes in raw material batches, equipment parameter drift, and sudden changes in environmental temperature and humidity, which makes the model unable to accurately quantify the uncertainty of prediction and the reliability of the prediction results is low. At this time, it is determined that the current process state exceeds the experience range and corresponding decision support information needs to be output.

[0042] It should be noted that the decision support information must be targeted and operable, and be able to guide operators to take conservative adjustment measures to avoid product quality problems caused by relying on unreliable prediction results; the content of the decision support information can be preset in advance according to common process abnormality scenarios to ensure the timeliness and effectiveness of the output.

[0043] In some possible embodiments, assuming that when the production of the aforementioned batch of photoresist is halfway complete, the viscosity of the raw material fluctuates significantly due to a temporary change in the batch by the raw material supplier, and this fluctuation characteristic does not appear in the model training data; at this time, the system monitors in real time that the width of the 95% dynamic confidence interval suddenly increases from 0.5μm to 0.8μm, which is greater than the width threshold of 0.6μm, and immediately determines that the current process state exceeds the empirical range; based on the preset abnormal scenario auxiliary information database, the system synchronously outputs the corresponding decision-making auxiliary information, including low prediction confidence and suggestions for conservative adjustments, such as reducing the raw material supply rate, increasing the stirring speed, and closely monitoring changes in the coating speed, to guide operators to take timely adjustment measures and reduce quality risks.

[0044] If the width of the obtained 95% dynamic confidence interval is less than or equal to the width threshold, the current process state is determined to be within the empirical range, and blank decision support information is output.

[0045] Specifically, when the width of the 95% dynamic confidence interval is less than or equal to the preset threshold, it indicates that the characteristics of the current process data are highly similar to the characteristics of the historical data on which the model training relies. The model can accurately capture the correlation between process parameters and quality indicators, with low prediction uncertainty and reliable prediction results. At this time, it is determined that the current process state is within the model's experience range, and there is no need to output additional decision support information, i.e., output blank decision support information. Operators can directly make subsequent decisions based on the predicted values ​​of quality indicators output by the model, such as whether to adjust process parameters or whether to continue to maintain the current production state, thereby reducing unnecessary intervention and improving production efficiency.

[0046] In some possible embodiments, during the production of a certain batch of photoresist, the width of the 95% dynamic confidence interval monitored in real time is always maintained between 0.4-0.5μm, which is less than or equal to the width threshold of 0.6μm. The system determines that the process status within this time period is within the empirical range and outputs blank decision support information. The operator does not need to adjust the process parameters, continues to maintain the current production status, and continuously monitors the subsequent prediction results.

[0047] The quality index prediction results are composed of the predicted values ​​of dry film photoresist quality indicators, the 95% dynamic confidence interval, and decision support information.

[0048] For example, continuing the above photoresist production process, the quality indicator prediction results can be integrated in two time periods. During the 9:00-11:20 time period, the predicted quality indicator value is film thickness uniformity ±2.6μm, with a 95% dynamic confidence interval of [±2.1μm, ±3.1μm], which is blank decision support information. The integrated prediction results are {film thickness uniformity ±2.6μm, [±2.1μm, ±3.1μm], none}. After 11:30, the predicted quality indicator value is film thickness uniformity ±2.9μm, with a 95% dynamic confidence interval of [±2.2μm, ±3.6μm]. The decision support information indicates low prediction confidence, suggesting conservative adjustments such as reducing the raw material supply rate, increasing the stirring speed, closely monitoring changes in coating speed, and integrating the predicted quality indicator results.

[0049] Step S3: Construct a staged few-sample network, and use the staged few-sample network to identify faults in the time-series process data and obtain fault diagnosis results.

[0050] like Figure 2 As shown, obtaining fault diagnosis results, i.e., step S3, in a data analysis method for the preparation of dry film photoresist includes the following steps: A staged few-shot network consisting of a first stage and a second stage is constructed and trained.

[0051] The first stage maps the time-series process data into real-time feature vectors based on the feature encoder, and the second stage performs fault type calibration on the time-series process data based on the feature encoder.

[0052] It should be noted that the aforementioned staged few-shot network reduces its dependence on fault samples through a two-stage training model. The first stage is normal pattern representation learning, and the second stage is few-shot prototype calibration. The two stages work together to ensure the network's feature extraction capability and fault identification accuracy. The first stage trains the feature encoder using massive amounts of unlabeled normal process data. Through a time-series comparison self-supervised learning task, the feature encoder learns the feature patterns of normal process data and has the ability to map time-series process data into high-dimensional real-time feature vectors. The two stages work together to ensure the network's feature extraction capability and fault identification accuracy. To avoid interference with normal pattern feature learning, the feature encoder parameters are frozen. A small number of labeled fault samples are used to calibrate the prototype vectors of each fault type through a metric learning method, enabling the network to distinguish the feature differences between different fault types.

[0053] The training process includes: Obtain normal process data and train the feature encoder based on the normal process data.

[0054] It should be noted that, in order to avoid insufficient learning of the feature encoder due to limited data and to ensure data integrity, it is necessary to acquire normal production data covering different operating conditions of the production line, such as normal process data under different raw material batches, different production loads, and different environmental conditions.

[0055] Specifically, the training process in this step adopts a self-supervised learning approach, eliminating the need for manual fault labeling and thus reducing data collection costs. Furthermore, during training, normal process data is preprocessed using the same method as in step S1, followed by data augmentation techniques such as time stretching and adding minor noise. The preprocessed data is then input into the feature encoder, and the model learns the inherent feature patterns of normal process data through a self-supervised task designed to correspond with production. Finally, the parameters of the feature encoder are optimized through iterative training until the model's feature discrimination on the validation set reaches a preset threshold, such as a feature similarity accuracy greater than or equal to 90%.

[0056] It should be noted that the design of the self-supervised tasks needs to be tailored to the characteristics of time-series data, guiding the model to learn the temporal correlation and trend features of process parameters. Simultaneously, the tasks must be clearly quantifiable to ensure a clear direction for model training optimization. Furthermore, the task difficulty must be appropriate to the model's capabilities to avoid convergence failure due to excessive difficulty. Common and reasonable self-supervised tasks are mainly divided into two categories: time-series comparison tasks and time-series reconstruction tasks. Time-series comparison tasks construct positive and negative sample pairs, allowing the model to learn to distinguish between similar and dissimilar time-series segments, thereby capturing the time-series patterns of normal processes. Time-series reconstruction tasks perform operations such as masking and shuffling on time-series data, allowing the model to reconstruct the original time series based on the remaining data, thereby learning the inherent dependencies of the time-series data.

[0057] Understandably, considering the precision requirements of photoresist production, to ensure the accurate capture of normal process data feature patterns by the feature encoder, and to effectively distinguish between normal process data and slightly abnormal process data, thus avoiding misjudgments in subsequent fault identification due to insufficient feature extraction, a threshold of 90% can be set. From an industrial practicality perspective, setting the threshold too high, such as 95% or above, will significantly increase the difficulty and time of model training, and may even lead to model overfitting, reducing its adaptability to normal data under new operating conditions. If the threshold is set too low, such as 85% or below, it will result in insufficient feature discrimination, blurring the feature boundaries between normal and abnormal data, leading to an increased misjudgment rate in subsequent fault identification. Therefore, considering both the model recognition accuracy requirements and the practicality of industrial scenarios, the threshold for feature similarity accuracy is set at 90%.

[0058] In some possible embodiments, assuming that when training the feature encoder corresponding to a certain batch of photoresist, time-series process data of 100 normal production batches in the history of the production line are collected, covering 3 types of raw material batches, 2 types of production loads (full load and half load), and 4 environmental temperature and humidity ranges; these data are preprocessed with cleaning and alignment consistent with step S1 to obtain standardized data; the length of each time series is stretched to 1.2 times the original length using a time stretching method to generate enhanced data; a time-series contrastive learning task is designed, dividing the time-series data of each batch into segments of 100 time points, generating 1 positive sample and 5 negative samples for each segment, where the positive sample represents adjacent samples of the same batch, and the negative sample represents samples of different batch segments; the processed data is input into the feature encoder, and the contrastive loss function is used to iteratively train it for 100 rounds using the Adam optimizer, and the finally trained feature encoder achieves a feature similarity accuracy of over 90% on the validation set.

[0059] The feature encoder is frozen, and fault process data with fault labels is obtained. The fault process data is input into the frozen feature encoder to generate fault sample feature vectors. The fault sample feature vectors are calibrated to obtain fault type vectors. A fault feature library is constructed based on the fault type vectors.

[0060] It should be noted that the operation of freezing the feature encoder is to preserve the normal process feature extraction capability learned in the first stage and avoid interference from the second stage training, thereby ensuring the accuracy of fault feature extraction. In addition, the fault process data with fault labels should include typical fault types, with 3-5 samples for each fault type, and the labels should accurately indicate the fault type and the time period of the fault occurrence.

[0061] Specifically, inputting fault process data into the frozen feature encoder generates fault sample feature vectors containing fault feature information; the calibration process integrates multiple feature vectors of the same fault type using statistical methods such as calculating the mean and median to obtain a fault type vector; finally, the prototype vectors of all fault types are integrated to construct a fault feature library. The fault feature library needs to be scalable so that new fault types can be added to it in the future.

[0062] In some possible embodiments, the feature encoder parameters trained in step S3-1-1 are frozen, and labeled fault process data of three typical fault types of the production line are collected, including coating speed drop fault (fault label: V1, fault characteristic: coating speed drops from 30mm / s to below 20mm / s within 10 seconds); drying temperature fluctuation fault (fault label: V2, fault characteristic: drying temperature fluctuates within ±5℃); and raw material supply pressure abnormal fault (fault label: V3, fault characteristic: raw material supply pressure is greater than 0.5MPa or less than 0.3MPa). After preprocessing these fault process data, they are input into the frozen feature encoder to generate a 64-dimensional fault sample feature vector for each sample. The feature vectors of the same fault type are calibrated, and the mean value of each dimension is calculated to obtain the fault type vector, including V1 (prototype vector of coating speed drop fault), V2 (prototype vector of drying temperature fluctuation fault), and V3 (prototype vector of raw material supply pressure abnormal fault). V1, V2, and V3 are integrated to construct a fault feature library.

[0063] A distance threshold is set. If the cosine distance between the real-time feature vector and the normal feature vector is less than or equal to the distance threshold, it is determined to be normal process data. If the cosine distance between the real-time feature vector and the normal feature vector is greater than the distance threshold, it enters the second stage.

[0064] It should be noted that the distance threshold is used to distinguish between normal process data and abnormal process data, and its value needs to be determined by statistical analysis of the feature vectors of normal process data. First, the cosine distance between the feature vectors of all historical normal process data and the normal feature vectors is obtained. The normal feature vector is represented by the mean of all normal feature vectors. With the goal of a false positive rate of less than or equal to 3% for normal data, the maximum cosine distance is determined as the distance threshold.

[0065] Understandably, if the misjudgment rate is too high, such as 5%, a large number of normal production states will be misjudged as abnormal, triggering unnecessary troubleshooting and process adjustments, leading to problems such as production interruption, decreased efficiency, and increased energy consumption. If the misjudgment rate is too low, such as 1%, although it can improve the accuracy of judgment, it will greatly increase the difficulty of setting the distance threshold, which may cause the model to be insensitive to minor anomalies and increase the risk of missed faults. Therefore, in the photoresist process, the distance threshold will be set with a normal data misjudgment rate of less than or equal to 3% as the target, which can reduce ineffective intervention and avoid missed faults.

[0066] Specifically, the current time-series process data is input into the feature encoder to generate a real-time feature vector. The cosine distance between the real-time feature vector and the normal feature vector is calculated. If the cosine distance is less than or equal to a set distance threshold, it indicates that the current process feature is highly similar to the normal process feature and is judged as normal process data. If the cosine distance is greater than the set distance threshold, it indicates that the current process feature deviates from the normal mode and there is an anomaly. Then, the second stage is entered for fault type matching.

[0067] In some possible embodiments, for a certain batch of photoresist production, the cosine distance between the feature vectors of 100 historical normal production batches and the normal feature vector is calculated. The maximum cosine distance is found to be 0.4. At this time, the misclassification rate of normal data is 2.8%, so the distance threshold is set to 0.4. When the production of this batch reaches 10:15, the real-time timing process data is input into the feature encoder to generate a 64-dimensional real-time feature vector V101. The cosine distance between V101 and the normal feature vector is calculated to be 0.3, which is less than or equal to the distance threshold of 0.4, and is judged as normal process data. When the production reaches 11:35, after the raw material batch is changed, a real-time feature vector V102 is generated. The cosine distance between it and the normal feature vector is calculated to be 0.5, which is greater than the distance threshold of 0.4, and is judged as abnormal process data, and enters the second stage for fault type matching.

[0068] Obtain the cosine distance between the real-time feature vector and the vectors of each fault type in the fault feature database, and filter out the minimum cosine distance.

[0069] Specifically, to determine the fault type corresponding to abnormal process data, the cosine distance between the real-time feature vector and each fault type vector in the fault feature library is calculated. The cosine distance quantifies the similarity between the real-time abnormal feature and each fault type feature. The smaller the value, the higher the similarity between the two vectors, that is, the more the real-time abnormal feature matches the fault type feature, and vice versa. The cosine distance between all real-time feature vectors and each fault type vector is counted, and the minimum value among all cosine distances is selected. The fault type corresponding to the minimum value is the most likely fault type of the current abnormality, providing a core basis for subsequent fault determination.

[0070] In some possible embodiments, assuming a real-time feature vector V102 is obtained, three fault type vectors V1, V2, and V3 are extracted from the fault feature library, and the cosine distance between V102 and each fault type vector is calculated, including a cosine distance of 0.6 between V102 and V1, a cosine distance of 0.7 between V102 and V2, and a cosine distance of 0.2 between V102 and V3; by comparison and selection, the smallest cosine distance of 0.2 is selected, and the corresponding fault type vector is V3, and it is preliminarily determined that the current anomaly is a raw material supply pressure abnormality fault.

[0071] A fault matching threshold is set. If the minimum cosine distance is less than or equal to the fault matching threshold, the fault type corresponding to the minimum cosine distance is taken as the fault diagnosis result.

[0072] Specifically, the fault matching threshold is determined based on the cosine distance of historical fault samples. First, the cosine distance between the feature vector of each historical fault sample and the corresponding fault type vector is calculated. With a fault identification accuracy of greater than or equal to 95% as the target, the largest cosine distance is determined as the fault matching threshold. If the smallest cosine distance is less than or equal to the threshold, it means that the similarity between the real-time abnormal feature and the corresponding fault type feature meets the identification requirements, and the current fault type can be clearly determined to be the fault type corresponding to the smallest cosine distance. If it is greater than the threshold, it means that the feature similarity between the current abnormal feature and all fault types in the fault feature library does not meet the identification requirements, and it may be a completely new fault type.

[0073] It should be noted that if the identification accuracy is too low, such as 90%, it is easy to misjudge the fault type. For example, an abnormal raw material supply pressure may be misjudged as a fluctuation in drying temperature, leading to incorrect direction of subsequent optimization instructions. This not only fails to solve the fault problem but may also exacerbate the process abnormality. When a fault occurs, it is necessary to quickly and accurately locate the type to reduce production interruption time. An accuracy of 95% can ensure the reliability of diagnosis while taking into account the efficiency of fault identification. It is important to avoid setting the threshold too strictly due to excessive pursuit of higher accuracy, which would increase the probability of missed faults and thus prolong the fault investigation time and reduce production efficiency. Therefore, in this embodiment, the fault identification accuracy is set at greater than or equal to 95%, which can avoid the risk of misjudgment and control the probability of missed faults.

[0074] In some possible embodiments, continuing the above example, the cosine distance between the feature vector of each historical fault sample and the corresponding fault type vector is statistically analyzed. It is found that when the cosine distance is less than or equal to 0.3, the fault identification accuracy reaches 96%, which meets the enterprise's requirement of identification accuracy greater than or equal to 95%. Therefore, the fault matching threshold is set to 0.3. According to the results of step S3-3, the minimum cosine distance is 0.2, which is less than or equal to the fault matching threshold of 0.3. Therefore, the fault type corresponding to the minimum cosine distance, raw material supply pressure abnormality fault, is taken as the fault diagnosis result.

[0075] Step S3-5: If the minimum cosine distance is greater than the fault matching threshold, the current fault is determined to be a completely new fault type, and the incremental learning mechanism is triggered.

[0076] Understandably, when the minimum cosine distance is greater than the fault matching threshold, it means that the feature pattern of the current abnormal process data has not appeared in the fault feature library and belongs to a completely new fault type. At this time, an incremental learning mechanism needs to be triggered to realize the rapid learning of the new fault type and the updating of the fault feature library.

[0077] It should be noted that the incremental learning mechanism calibrates new fault type vectors using a small number of brand-new fault samples without retraining the entire staged small sample network, and adds them to the fault feature library, ensuring that the network's fault identification capability can be continuously expanded, thereby adapting to the new fault scenarios that constantly emerge in industrial production.

[0078] Furthermore, the execution steps of the incremental learning mechanism include automatic acquisition: after the system triggers the mechanism, it automatically acquires sufficient time-series process data samples within the time period of the new fault occurrence to ensure that the samples can completely represent the feature patterns of the new fault; feature extraction: the acquired new fault samples are input into the frozen feature encoder to generate corresponding fault sample feature vectors; vector calibration: by calculating the mean or median of multiple sample feature vectors under the same new fault type, a new fault type vector that can represent the fault feature is obtained; library update: the new fault type vector, the corresponding manually labeled fault tags, such as coating roller temperature abnormality fault, and fault feature descriptions, such as coating roller temperature rising sharply by 10°C within 5 seconds, are added to the fault feature library to complete the incremental update.

[0079] For example, suppose that during the production of a certain batch of photoresist, the cosine distance between the real-time feature vector V103 and the normal feature vector is 0.5, which is greater than the distance threshold of 0.4. After entering the second stage, the cosine distances between V103 and V1, V2, and V3 in the fault feature library are calculated to be 0.5, 0.6, and 0.4, respectively. The minimum cosine distance is 0.4, which is greater than the fault matching threshold of 0.3. The system determines that the current fault is a completely new fault type and immediately triggers the incremental learning mechanism. First, it automatically collects five time-series process data samples of the new fault. Then, it inputs the samples into the frozen feature encoder to generate feature vectors. Next, it calculates the mean of the feature vectors to obtain the new fault type vector V4. Finally, it adds V4 and the corresponding fault label to the fault feature library, thereby completing the incremental learning.

[0080] Step S4: Based on the quality index prediction results and fault diagnosis results, and combined with the hierarchical reinforcement learning framework, process parameter optimization instructions are generated.

[0081] Step S4 includes the following: Define the process regulation subtask, introduce the SAC algorithm into the high-level policy controller and the low-level policy controller, and build a hierarchical reinforcement learning framework based on the high-level policy controller and the low-level policy controller.

[0082] Specifically, by combining the process knowledge of dry film photoresist preparation, process adjustment subtasks are defined, decomposing the complex overall optimization objective into multiple subtasks with clear process semantics, such as rapidly stabilizing the coating speed and finely adjusting the temperature difference in the drying zone. Each subtask corresponds to a specific type of process adjustment requirement. The SAC algorithm is introduced into the high-level and low-level policy controllers to handle the continuous actions of fine-tuning process parameters. Specifically, the high-level controller is responsible for selecting tasks, i.e., choosing appropriate process options, while the low-level controller is responsible for performing actions, i.e., executing the specific parameter fine-tuning of the selected process options. Through the collaborative work of the two layers of controllers, the complexity of the decision space is reduced, thereby improving the efficiency and accuracy of optimization decisions.

[0083] It should be noted that the SAC algorithm is a deep reinforcement learning algorithm based on maximum entropy reinforcement learning. It consists of an actor network, a critic network, and a target network. By introducing an entropy regularization term, a maximum entropy objective function is constructed to achieve policy optimization while stabilizing the exploration process. Specifically, the actor network is responsible for generating continuous actions that conform to the current process state, such as specific adjustment values ​​of process parameters; the critic network is responsible for evaluating the value of the actions, i.e., the probability of achieving the quality standard after executing the action; the target network is used to stabilize the training process and avoid training oscillations; the entropy regularization term encourages the policy to explore more potential high-quality actions while preventing the policy from converging to a local optimum too early, ensuring adaptability in complex process scenarios.

[0084] Understandably, this step adapts to the continuous action space by introducing the SAC algorithm, thus meeting the needs of process fine-tuning. The process parameters for dry film photoresist preparation are all continuous values, such as supply pressure, stirring speed, and coating speed. Traditional discrete action algorithms are difficult to achieve fine-tuning, while the SAC algorithm supports continuous action output and can accurately calculate parameter adjustment values ​​that meet equipment constraints, avoiding process fluctuations caused by parameter mutations. At the same time, the dry film photoresist preparation process is easily affected by various factors such as raw material batches and environmental temperature and humidity, and the process state changes dynamically. The entropy regularization mechanism of the SAC algorithm can balance the exploration of new actions with the utilization of existing optimal actions, and can quickly adapt to process changes such as raw material changes and equipment parameter drift, ensuring the stability of optimization decisions.

[0085] In some possible embodiments, continuing the above example, four process adjustment sub-tasks are defined based on process knowledge, including A. rapidly stabilizing the coating speed, B. finely adjusting the temperature difference in the drying zone, C. stabilizing the raw material supply pressure, and D. enhancing the uniformity of raw material mixing. The SAC algorithm is introduced to construct a high-level strategy controller and a low-level strategy controller. The high-level controller takes the quality index prediction result, fault diagnosis result, and current process status as inputs, which can be obtained based on the quality index prediction result, and outputs a process option selection command. The low-level controller takes the process option command output by the high-level controller and the current process parameters as inputs, and outputs specific optimized values ​​for the process parameters.

[0086] Furthermore, when training the hierarchical reinforcement learning framework, a combination of historical production data and simulation data can be used. By setting reward functions to optimize controller parameters, such as rewards for improving quality compliance rate and rewards for reducing energy consumption, a hierarchical reinforcement learning framework that can adapt to different process conditions can be finally constructed.

[0087] Based on the process adjustment subtask, the quality indicator prediction results and fault diagnosis results are input into the high-level strategy controller, and process option instructions are output.

[0088] Specifically, the high-level controller uses the policy network of the SAC algorithm to perform feature fusion and probabilistic reasoning on the input information, evaluate the implementation value of each process adjustment subtask, that is, the probability of achieving the quality indicators and energy consumption cost after executing the subtask, and select the process adjustment subtask with the highest implementation value as the output process option instruction.

[0089] For example, the predicted results of quality indicators (film thickness uniformity ±2.9μm, 95% confidence interval [±2.2μm, ±3.6μm], low confidence of decision support-information prediction, conservative adjustment recommended) and the fault diagnosis results (abnormal raw material supply pressure fault) are input into the high-level strategy controller. The high-level controller evaluates the implementation value of four process options through the SAC algorithm strategy network, including option A (stabilize coating speed) implementation value 0.3, option B (adjust drying temperature difference) implementation value 0.2, option C (stabilize raw material supply pressure) implementation value 0.8, and option D (enhance mixing uniformity) implementation value 0.4. Option C, which has the highest implementation value, is selected as the process option instruction, and the output is to select process option C to stabilize the raw material supply pressure.

[0090] The process option instructions are input into the underlying strategy controller, and process parameter optimization instructions are output.

[0091] Specifically, the underlying controller uses the action network of the SAC algorithm to calculate the optimal process parameter adjustment value by combining the adjustment target of the selected process option and the actual value of the current process parameter. At the same time, it considers the adjustment constraints of the process parameters, such as the adjustment range not exceeding the allowable range of the equipment and the adjustment rate not being too fast to avoid equipment damage. The final output process parameter optimization instruction must include information such as the name of the adjustment parameter, the current value, the target value, and the adjustment rate.

[0092] In some possible embodiments, it is assumed that the instruction to select process option C, stabilize the raw material supply pressure, output by the high-level controller is input into the low-level strategy controller. The low-level controller, based on the SAC algorithm action network, combines the adjustment target of option C, such as adjusting the pressure to about 0.4 MPa, and the equipment adjustment constraints, with a maximum adjustment rate of 0.02 MPa / second for the raw material supply pressure, to calculate the process parameter optimization instruction, including the current pressure value of 0.55 MPa, the target value of 0.4 MPa, and the adjustment rate of 0.02 MPa / second. After the adjustment is completed, the pressure value is maintained, and the pressure data is monitored every 10 seconds. This instruction is directly sent to the PLC system of the production line to automatically execute the raw material supply pressure adjustment operation.

[0093] Figure 3 The diagram illustrates a data analysis system for the preparation of dry film photoresist, provided by some embodiments of this application, which can realize the ideas of this application.

[0094] Specifically, a data analysis system for the preparation of dry film photoresist includes: The acquisition module is used to acquire the timing process data of the entire dry film photoresist preparation process; The processing module is used to preprocess the time-series process data, generate process parameter optimization instructions based on the quality index prediction results and fault diagnosis results, and in combination with the hierarchical reinforcement learning framework. The input module inputs the timing process data into the TCN-NP model; The output module outputs the quality index prediction results of the dry film photoresist based on the TCN-NP model; The building module is used to build a staged few-shot network; The identification module identifies faults in time-series process data based on a staged few-sample network and obtains fault diagnosis results.

[0095] The specific usage and function of this embodiment are explained below: First, the timing process data of the entire dry film photoresist fabrication process is acquired and preprocessed. Then, the timing process data is input into the TCN-NP model, which outputs the quality index prediction results of the dry film photoresist. Next, a staged few-sample network is constructed to identify faults in the timing process data and obtain fault diagnosis results. Finally, based on the quality index prediction results and fault diagnosis results, and combined with a hierarchical reinforcement learning framework, process parameter optimization instructions are generated. This integrated technical solution, which enables accurate processing of the entire process timing data, quality prediction and uncertainty quantification, few-sample fault identification, and collaborative optimization decision-making, solves the problems of low process control accuracy and blind optimization decisions in the dry film photoresist fabrication process, thereby improving the stability of photoresist fabrication quality.

[0096] This embodiment provides an electronic device, which may include: at least one processor, at least one network interface, a user interface, a memory, and at least one communication bus.

[0097] The following is a detailed introduction to the various components of the electronic device: The communication bus can be used to enable communication between the various components mentioned above.

[0098] The user interface may include buttons, and optional user interfaces may also include standard wired interfaces and wireless interfaces.

[0099] The network interface may include, but is not limited to, Bluetooth modules, NFC modules, Wi-Fi modules, etc.

[0100] The processor may include one or more processing cores. It connects various parts of the electronic device via various interfaces and lines, executing instructions, programs, code sets, or instruction sets stored in memory, and accessing data stored in memory to perform various functions and process data. Optionally, the processor can be implemented using at least one hardware form of DSP, FPGA, or PLA. The processor can integrate one or more of the following: CPU, GPU, and modem, for example, one or more digital signal processors (DSPs) or one or more field-programmable gate arrays (FPGAs). The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor.

[0101] The memory may include RAM or ROM. Optionally, the memory may include a non-transitory computer-readable medium. The memory may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (e.g., touch function, sound playback function, image playback function, etc.), instructions for implementing the various method embodiments described above, etc.; the data storage area may store data involved in the various method embodiments described above, etc. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor. The memory, as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an evaluation application. The processor may be used to call the evaluation application stored in the memory and execute the method steps mentioned in the foregoing embodiments.

[0102] It should be noted that the above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0103] The above embodiments can be implemented, in whole or in part, through software, hardware (such as circuits), firmware, or any other combination thereof.

[0104] When implemented using software, the above embodiments can be implemented in whole or in part as a computer program product, which includes one or more computer instructions or computer programs; when the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in accordance with the embodiments of the present invention are generated in whole or in part.

[0105] It is understood that the computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device; the computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via transmission methods such as infrared, wireless, or microwave; the computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0106] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0107] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0108] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for analyzing data from the preparation of dry film photoresist, characterized in that, It includes the following steps: Acquire the timing process data of the entire dry film photoresist preparation process, and preprocess the timing process data; The time-series process data is input into the TCN-NP model, and the quality index prediction results of dry film photoresist are output. The TCN-NP model includes a feature extraction layer and a probabilistic inference layer. The time-series process data is input into the feature extraction layer of the TCN-NP model, and a high-dimensional time-series feature vector is extracted through a dilated causal convolution structure. Then, the probabilistic inference layer outputs the quality index prediction value of dry film photoresist, 95% dynamic confidence interval and decision support information, which together form the quality index prediction results. A staged few-sample network is constructed, and fault identification is performed on time-series process data based on the staged few-sample network to obtain fault diagnosis results. An incremental learning mechanism is triggered for new fault types. Based on the quality index prediction results and fault diagnosis results, and combined with the hierarchical reinforcement learning framework to generate process parameter optimization instructions, the hierarchical reinforcement learning framework introduces the SAC algorithm to the high-level policy controller and the low-level policy controller. The high-level policy controller outputs process option instructions, and the low-level policy controller outputs process parameter optimization instructions.

2. The method for analyzing data from the preparation of dry film photoresist according to claim 1, characterized in that, Acquire timing process data for the entire dry film photoresist fabrication process, and preprocess the timing process data, including: Based on distributed acquisition, raw material supply parameters, stirring parameters, coating speed, drying temperature, curing time, and ambient temperature and humidity are obtained and combined to form the time-series process data of the entire dry film photoresist preparation process. The time-series process data is preprocessed by cleaning and alignment to obtain standardized time-series process data. The cleaning includes removing outliers and filling in missing values ​​in the time-series process data. The alignment refers to unifying the time scale of the time-series process data in all dimensions.

3. The method for analyzing data from the preparation of dry film photoresist according to claim 1, characterized in that, Input the timing process data into the TCN-NP model, and output the predicted quality indicators of the dry film photoresist, including: A TCN-NP model is constructed based on temporal convolutional networks and neural processes. The TCN-NP model includes a feature extraction layer and a probabilistic inference layer. Among them, temporal convolutional networks are used as deterministic feature extractors to replace multilayer perceptrons in neural processes; Time-series process data is input into the feature extraction layer, which extracts high-dimensional time-series feature vectors based on a dilated causal convolution structure. The high-dimensional temporal feature vector is input into the probabilistic inference layer. Probability calculation is performed based on the latent variable framework of the probabilistic inference layer, and the predicted value of the dry film photoresist quality index and the 95% dynamic confidence interval are output. Assisted decision-making is based on a 95% dynamic confidence interval to obtain decision support information; The quality index prediction results are composed of the predicted values ​​of dry film photoresist quality indicators, the 95% dynamic confidence interval, and decision support information.

4. The method for analyzing data from the preparation of dry film photoresist according to claim 3, characterized in that, Decision support is based on a 95% dynamic confidence interval to obtain decision support information, including: Set a width threshold and monitor the width of the 95% dynamic confidence interval in real time; If the width of the obtained 95% dynamic confidence interval is greater than the width threshold, it is determined that the current process state exceeds the empirical range, and the corresponding decision support information is output simultaneously. If the width of the obtained 95% dynamic confidence interval is less than or equal to the width threshold, the current process state is determined to be within the empirical range, and blank decision support information is output.

5. The method for analyzing data from the preparation of dry film photoresist according to claim 1, characterized in that, A staged few-sample network is constructed, and fault identification is performed on time-series process data based on the staged few-sample network to obtain fault diagnosis results, including: Construct a staged few-shot network comprising a first stage and a second stage, and train the staged few-shot network. The first stage maps the time-series process data into real-time feature vectors based on the feature encoder, and the second stage performs fault type calibration on the time-series process data based on the feature encoder. Set a distance threshold. If the cosine distance between the real-time feature vector and the normal feature vector is less than or equal to the distance threshold, it is determined to be normal process data. If the cosine distance between the real-time feature vector and the normal feature vector is greater than the distance threshold, it enters the second stage. Obtain the cosine distance between the real-time feature vector and the vectors of each fault type in the fault feature library, and filter out the minimum cosine distance; Set a fault matching threshold. If the minimum cosine distance is less than or equal to the fault matching threshold, then the fault type corresponding to the minimum cosine distance is taken as the fault diagnosis result. If the minimum cosine distance is greater than the fault matching threshold, the current fault is determined to be a completely new fault type, and an incremental learning mechanism is triggered.

6. The method for analyzing data from the preparation of dry film photoresist according to claim 5, characterized in that, Training a staged few-shot network includes: Obtain normal process data and train the feature encoder based on the normal process data; The feature encoder is frozen, and fault process data with fault labels is obtained. The fault process data is input into the frozen feature encoder to generate fault sample feature vectors. The fault sample feature vectors are calibrated to obtain fault type vectors. A fault feature library is constructed based on the fault type vectors.

7. The method for analyzing data from the preparation of dry film photoresist according to claim 1, characterized in that, Based on the predicted results of quality indicators and the results of fault diagnosis, and combined with a hierarchical reinforcement learning framework, process parameter optimization instructions are generated, including: Define the process regulation subtask, introduce the SAC algorithm to the high-level policy controller and the low-level policy controller, and build a hierarchical reinforcement learning framework based on the high-level policy controller and the low-level policy controller. Based on the process adjustment subtask, the quality indicator prediction results and fault diagnosis results are input into the high-level strategy controller, and the process option instructions are output. The process option instructions are input into the underlying strategy controller, and process parameter optimization instructions are output.

8. A data analysis system for the preparation of dry film photoresist, used to implement the method described in any one of claims 1-7, characterized in that, include: The acquisition module is used to acquire the timing process data of the entire dry film photoresist preparation process; The processing module is used to preprocess the time-series process data, generate process parameter optimization instructions based on the quality index prediction results and fault diagnosis results, and in combination with the hierarchical reinforcement learning framework. The input module inputs the timing process data into the TCN-NP model; The output module outputs the quality index prediction results of the dry film photoresist based on the TCN-NP model; The building module is used to build a staged few-shot network; The identification module identifies faults in time-series process data based on a staged few-sample network and obtains fault diagnosis results.

Images