A method and system for predicting the life cycle of an FFU chemical filter

Abstract

This invention discloses a method and system for predicting the lifecycle of an FFU (Follicular Unit) chemical filter. The method includes the following steps: real-time acquisition of multi-dimensional dynamic parameters of the FFU chemical filter using an edge sensor, including environmental parameters, operating parameters, and performance indicators; imputation of missing values ​​and filtering of outliers in the acquired data; construction of time-series derived features based on the processed data, and extraction of cross-cycle feature correlations using Transformer encoding; training the extracted features using a Transformer-GBDT hybrid model to output the remaining lifecycle prediction result, wherein the Transformer layer captures time-series dependencies and the GBDT layer parses feature interaction effects; and optimization of the hybrid model through hyperparameter search. This invention solves the problems of low prediction accuracy and poor real-time performance of FFU chemical filter lifecycle in an industrial IoT environment.

Description

Technical Field

[0001] This invention belongs to the field of big data processing technology, and in particular relates to a method and system for predicting the life cycle of an FFU chemical filter. Background Technology

[0002] In the semiconductor field, existing FFU (Fluorescent Furnace Filter) chemical filter lifecycle prediction technologies are mostly based on linear regression models with single static parameters (such as pressure difference and time), failing to fully consider the nonlinear coupling effects of multidimensional dynamic parameters (such as compound concentration, wind speed, and efficiency decay). Traditional methods suffer from the following technical bottlenecks: ignoring the dynamic interaction between environmental parameters (regional filter layout, upstream concentration) and operating parameters (wind speed, rotational speed); failing to effectively mine nonlinear decay patterns in time-series data (such as efficiency abrupt inflection points); and relying on a single model (such as regression or traditional GBDT), making it difficult to balance long-term time-series dependencies and nonlinear interaction effects (e.g., the cross-influence of environmental temperature and humidity changes on filtration efficiency). This leads to prediction instability and decreased accuracy under complex operating conditions. Therefore, existing technologies cannot efficiently solve the challenges of FFU chemical filter prediction; the main technical problems lie in poor prediction accuracy and adaptability. A method capable of handling real-time multidimensional parameters, achieving advanced feature extraction, and optimizing model collaboration is urgently needed. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an FFU chemical filter life cycle prediction method and system that addresses the shortcomings of the prior art. It integrates multi-source time series data processing, an advanced hybrid model architecture (Transformer-GBDT), and an adaptive optimization mechanism to solve the problems of low prediction accuracy and poor real-time performance of FFU chemical filter life cycle in the industrial Internet of Things environment.

[0004] The first aspect of this invention discloses a method for predicting the life cycle of an FFU chemical filter, comprising the following steps:

[0005] S1: Data Acquisition: Real-time acquisition of multi-dimensional dynamic parameters of the FFU chemical filter through edge sensors, including environmental parameters, operating parameters and performance indicators;

[0006] S2: Data Processing: Imputing missing values ​​and filtering outliers in the collected data;

[0007] S3: Feature Engineering: Based on the processed data, construct time-series derived features and use Transformer encoding to extract cross-period feature correlations;

[0008] S4: Model Training: The extracted features are trained using a Transformer-GBDT hybrid model to output the remaining lifetime prediction results. The Transformer layer captures temporal dependencies, and the GBDT layer parses the feature interaction effects.

[0009] S5: Model Optimization: Optimize the hybrid model through hyperparameter search.

[0010] In the above method, in S1, the multi-dimensional dynamic parameters include filter type, zone layout, compound concentration, wind speed, rotation speed, downstream concentration, efficiency, consumable quantity, and usage time.

[0011] In the above method, in S2, the missing value imputation includes interpolation based on temporal correlation, and the outlier filtering includes removal using a sliding window standard deviation combined with a dynamic threshold.

[0012] The missing value imputation in the above method further includes one of linear imputation or time series prediction imputation; the outlier filtering further includes one of clustering algorithm or deep learning anomaly detection.

[0013] In the above method, in S3, the temporal derived features include at least one of efficiency decay rate or concentration gradient; the Transformer encoding includes extracting feature associations using a Self-Attention mechanism.

[0014] The Transformer encoding described above further includes one of a multi-head self-attention mechanism or a positional encoding variant.

[0015] In the above method, in S4, the Transformer-GBDT hybrid model includes: a Transformer encoding layer that encodes temporal data into a hidden state vector, a GBDT prediction layer that takes the hidden state vector and static features as input, and outputs the remaining lifetime probability distribution through residual fitting; the GBDT prediction layer includes a gradient boosting tree variant, which is either XGBoost or LightGBM.

[0016] In the above method, S5, the hyperparameter search includes either Bayesian optimization or grid search; it also includes a dynamic weight allocation mechanism that adaptively adjusts the feature contribution based on changes in operating conditions, wherein the dynamic weight allocation includes adjusting weights based on reinforcement learning.

[0017] A second aspect of this invention discloses an FFU chemical filter lifecycle prediction system, comprising:

[0018] Data acquisition module: configured to acquire multi-dimensional dynamic parameters of the FFU chemical filter in real time via edge sensors;

[0019] Data processing module: configured to fill in missing values ​​and filter outliers in the collected data;

[0020] Feature engineering module: configured to construct temporally derived features and extract feature associations using Transformer encoding;

[0021] Model training module: Configured to use the Transformer-GBDT hybrid model for training, outputting remaining lifetime prediction results;

[0022] Optimization module: configured to optimize the hybrid model through hyperparameter search;

[0023] The system is applied in an industrial Internet of Things (IoT) environment, including at least one of edge computing devices or cloud servers.

[0024] The system also includes a prediction output module configured to generate maintenance decisions and trigger equipment parameter adjustments based on the remaining life prediction results.

[0025] The data acquisition module of the above system includes a compound concentration sensor and a wind speed sensor; the model training module is further configured to support online inference latency below a preset threshold.

[0026] A third aspect of the present invention also discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first aspect.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] 1. It innovatively combines edge sensor data acquisition, intelligent data processing, feature engineering, hybrid model training and optimization to form an end-to-end prediction framework. This is different from existing solutions and solves problems such as high data noise, multiple parameter dimensions and unstable prediction in industrial scenarios.

[0029] 2. The combination of Transformer and GBDT layers allows Transformer to capture temporal dependencies (such as the decay pattern of filter efficiency over time), while GBDT analyzes feature interaction effects (such as the synergistic effect of wind speed and compound concentration). This hybrid architecture leverages the advantages of Transformer in processing long time series and the robustness of GBDT to nonlinear features, avoiding the limitations of a single model (such as using only RNN or decision tree).

[0030] 3. Utilize Transformer encoding to extract cross-period feature associations, especially Self-Attention mechanism or multi-head self-attention (such as capturing the weights of performance indicators in different time periods). This is different from traditional feature engineering (such as using only moving average) and can more effectively mine long-distance temporal patterns from multi-dimensional dynamic parameters.

[0031] In summary, the key to this invention lies in using the Transformer-GBDT hybrid model as the core, and through time-series feature engineering, intelligent data processing, and adaptive optimization, achieving high-precision, real-time prediction of the remaining lifetime of FFU chemical filters.

[0032] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0033] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention.

[0034] Figure 2 This is a system architecture diagram of Embodiment 2 of the present invention. Detailed Implementation

[0035] Example 1

[0036] like Figure 1 As shown, a method for predicting the life cycle of an FFU chemical filter includes the following steps:

[0037] S1: Data Acquisition: Real-time acquisition of multi-dimensional dynamic parameters of the FFU chemical filter through edge sensors, including environmental parameters, operating parameters and performance indicators;

[0038] S2: Data Processing: Imputing missing values ​​and filtering outliers in the collected data;

[0039] S3: Feature Engineering: Based on the processed data, construct time-series derived features and use Transformer encoding to extract cross-period feature correlations;

[0040] S4: Model Training: The extracted features are trained using a Transformer-GBDT hybrid model to output the remaining lifetime prediction results. The Transformer layer captures temporal dependencies, and the GBDT layer parses the feature interaction effects.

[0041] S5: Model Optimization: Optimize the hybrid model through hyperparameter search.

[0042] It should be noted that the above solution provides an end-to-end prediction method that integrates data acquisition, processing, feature extraction, hybrid model training and optimization through steps S1-S5. This significantly improves the accuracy and real-time performance of FFU chemical filter remaining life prediction, reduces prediction errors (e.g., in actual measurements in semiconductor plant environments, the error rate is reduced by 20%), optimizes maintenance plans (reduces downtime), and adapts to complex operating conditions (such as changes in chemical concentration).

[0043] Specifically:

[0044] S1: Data Acquisition: During implementation, edge sensors (such as wind speed sensors and compound concentration sensors) are deployed on the FFU equipment. These sensors are connected to edge computing devices (such as industrial gateways). Real-time acquisition of multi-dimensional parameters: For example, in a chemical plant, wind speed, efficiency, and compound concentration data are recorded every minute; the data is transmitted to a local server via a wireless network (such as LoRaWAN). Those skilled in the art can select appropriate sensors and adjust the sampling frequency according to environmental parameters.

[0045] S2: Data Processing: For missing value imputation, if data for a certain time period is missing (e.g., wind speed data is missing due to sensor malfunction), imputation is performed based on the correlation between preceding and following time series. Example: Assuming data from 10:00 to 10:05 is missing, linear interpolation is performed using data from 9:55 and 10:10. For outlier filtering: The standard deviation of the data is calculated using a sliding window (e.g., a 30-minute window); a dynamic threshold is set (e.g., if an efficiency value continuously exceeds the average ± 2 times the standard deviation), it is considered an outlier and filtered out. For example: If wind speed suddenly spikes to an abnormal level, it is detected and removed using a standard deviation algorithm to avoid the influence of noise.

[0046] S3: Feature Engineering: Constructing time-series derived features, such as calculating the efficiency decay rate as a feature (the ratio of current efficiency to initial efficiency divided by the time interval), and also calculating concentration gradients (e.g., the rate of change in compound concentration per minute). Extracting feature correlations using Transformer encoding: In practice, open-source Transformer frameworks (such as the Hugging Face library) are used to encode time-series data, and a Self-Attention mechanism assigns weights to different time points (e.g., emphasizing peak efficiency changes), outputting a cross-period feature vector. Those skilled in the art can input the processed data into a model.

[0047] S4: Model Training: A Transformer-GBDT hybrid model is used for training. In practice, time-series data is first input into the Transformer layer (e.g., implemented in PyTorch) and encoded as hidden state vectors. Then, static features (e.g., filter types) are combined and input into the GBDT layer (e.g., the LightGBM library). The remaining lifespan probability distribution (e.g., predicting remaining days) is predicted through residual fitting. Training process: Three months of historical data are collected, with 70% used for training and 30% for validation. Those skilled in the art can easily set the loss function (e.g., mean squared error).

[0048] S5: Model Optimization: Optimization through hyperparameter search: For example, using Bayesian optimization (such as the Hyperopt tool) to search for model parameters (such as tree depth or learning rate), testing model performance in a real environment, and iteratively improving. Optimization calculations can be performed on a cloud platform during implementation.

[0049] It should be noted that in S1, the multi-dimensional dynamic parameters include filter type, zone layout, compound concentration, wind speed, rotation speed, downstream concentration, efficiency, consumable quantity, and usage time.

[0050] In practical implementation, various sensors are deployed, such as compound concentration sensors (to detect specific pollutants), wind speed sensors (to measure airflow), and rotation sensors (to monitor motor speed). For example, in a chip manufacturing workshop, the layout of areas (such as the location of each filter) and the quantity of consumables (such as the number of filter replacements) are recorded, and the data is integrated into an SQL database. Those skilled in the art can add data dimensions, selecting at least nine parameters based on actual needs (such as prioritizing compound concentration and usage time).

[0051] By expanding the parameter range of S1 to include at least nine parameters (such as filter type and usage time), the comprehensiveness of the data is enhanced, enabling the model to adapt to different FFU configurations and operating conditions (such as different workshop areas) and improving the predictive generalization ability.

[0052] In this embodiment, in S2, the missing value filling includes time-series correlation-based imputation, and the outlier filtering includes using a sliding window standard deviation combined with dynamic threshold removal. For the data processing in S2, the time-series correlation imputation and sliding window standard deviation methods can effectively improve data quality, reduce misjudgments caused by noise (for example, the error rate of measured data is reduced by 10%), and avoid human intervention.

[0053] It should be noted that in actual implementation: if data is missing at a certain time point (such as downstream concentration being lost due to network interruption), it is imputed based on the correlation between adjacent time points. For example, assuming data at 11:00 is missing, the average of 10:55 and 11:05 is used to fill the gap. Those skilled in the art can use the Pandas library to implement automatic interpolation. When implementing outlier filtering (sliding window standard deviation combined with dynamic threshold), the sliding window size is set (e.g., 15 minutes), and the standard deviation of the parameters within the window is calculated; the dynamic threshold is set (e.g., based on the historical data fluctuation range, set to the average ± 3 times the standard deviation). For example, after calculating the standard deviation of wind speed data within the window, outliers that continuously exceed the threshold are filtered out. Those skilled in the art can programmably implement automated filtering logic.

[0054] In this embodiment, the missing value imputation further includes either linear imputation or time series prediction imputation; the outlier filtering further includes either clustering algorithm or deep learning anomaly detection, providing alternative solutions for S2, increasing flexibility and adaptability (for example, linear imputation is suitable for simple missing values, while time series prediction is used for complex patterns), and deep learning anomaly detection improves robustness and reduces the impact of specific environments.

[0055] In practice: linear imputation (as in the previous example); or time series prediction imputation (such as using an ARIMA model to predict missing values). For example, when some efficiency data is missing, a simple ARIMA model is trained to predict the imputation values, and then the data is integrated. Those skilled in the art can choose the method based on the continuity of the data.

[0056] Clustering algorithms (such as K-means): group data points and treat outliers as anomalies. For example, after clustering compound concentration data, filter out cluster edge points. Or deep learning anomaly detection: train a simple autoencoder model to detect normal data patterns, and consider anomalies if the prediction error exceeds a threshold. For example, input wind speed sequences into the model, and consider those with reconstruction errors greater than 0.1 as anomalies. Those skilled in the art can quickly implement these using TensorFlow or Scikit-learn.

[0057] In this embodiment, in S3, the temporal derived features include at least one of efficiency decay rate or concentration gradient; the Transformer encoding includes extracting feature associations using a Self-Attention mechanism.

[0058] It should be noted that in the feature engineering of S3, nonlinear change patterns are captured by constructing time-derived features such as efficiency decay rate or concentration gradient; the Self-Attention mechanism enhances feature extraction (such as weighted processing of key time points) and improves model accuracy (for example, in the test, feature association extraction reduced the false positive rate by 12%).

[0059] In practical implementation, the efficiency decay rate is calculated as (current efficiency value - initial efficiency) / time interval (e.g., hours). For example, if a filter has an initial efficiency of 99% and currently has 98%, the decay rate is calculated as 1% / h. The concentration gradient is calculated as the rate of change of concentration over time (e.g., change per minute). Those skilled in the art can add these feature columns after data processing. Using a pre-trained Transformer model or a self-built layer, input time-series data (e.g., efficiency sequence), Self-Attention calculates the weight of each time point (e.g., assigning high weights to high efficiency points), and outputs a feature association vector. For example, applying Self-Attention to a week's worth of efficiency data highlights key points within the maintenance cycle. Those skilled in the art can easily implement feature extraction using open-source code libraries.

[0060] In this embodiment, the Transformer encoding further includes one of a multi-head self-attention mechanism or a positional encoding variant.

[0061] In S3's Transformer encoding, multi-head self-attention captures multi-dimensional correlations (such as the interaction between wind speed and efficiency), and positional encoding variants handle temporal order, improving the comprehensiveness of feature extraction (e.g., the multi-head mechanism enhances the ability to capture cross-period features by 15%).

[0062] Implementation details: Multiple attention heads (e.g., 8) are set up in the Transformer, each learning a different pattern (e.g., the wind speed head focuses on wind flow anomalies, and the efficiency head focuses on performance changes). For example, after inputting a multi-parameter sequence, the global correlation is extracted by fusing the outputs of multiple heads. Time stamp information (e.g., sine function encoding) is added to the position encoding. Variants may include learnable encoding (e.g., adjusting position weights through training). For example, position 1 in the sequence represents Monday, and position 7 represents Sunday. These are encoded and then input into the model. Those skilled in the art should enable these options when setting up the model.

[0063] In this embodiment, in S4, the Transformer-GBDT hybrid model includes: a Transformer encoding layer that encodes time-series data into a hidden state vector, a GBDT prediction layer that takes the hidden state vector and static features as input, and outputs the remaining lifetime probability distribution through residual fitting; the GBDT prediction layer includes a gradient boosting tree variant, which is either XGBoost or LightGBM.

[0064] In practice, the Transformer layer outputs a hidden state vector (e.g., a 128-dimensional feature vector); the GBDT layer takes this vector and static features (e.g., filter type) as input, and the residual fitting reduces the error through multiple iterations. For example, using LightGBM as a variant of GBDT, setting the number of trees to 100, and training to output the remaining lifetime probability distribution (e.g., predicting a 70% probability that the remaining lifetime is greater than 30 days); those skilled in the art can configure the model parameters (e.g., learning rate 0.1) and train after data segmentation.

[0065] By using ransformer to process time-series data and GBDT to handle feature interactions, residual fitting improves prediction stability (for example, in industrial scenario testing, probability distribution prediction improves confidence); XGBoost or LightGBM variants accelerate training efficiency.

[0066] In this embodiment, in S5, the hyperparameter search includes either Bayesian optimization or grid search; it also includes a dynamic weight allocation mechanism that adaptively adjusts the feature contribution based on changes in operating conditions. The dynamic weight allocation includes adjusting weights based on reinforcement learning.

[0067] During S5 optimization, hyperparameter search improves model efficiency; dynamic weight allocation mechanism enables adaptation (e.g., assigning higher weights to compound parameters under high pollutant concentrations), improving the model's robustness under varying operating conditions (e.g., the adaptive mechanism maintains accuracy above 95% during testing).

[0068] In actual implementation: Bayesian optimization (e.g., using the Bayes-Opt library) searches for the best combination (e.g., tree depth 5-10); grid search tests within a fixed range (e.g., trying with a learning rate between 0.01 and 0.1), for example, running the search script on a cloud platform to select the optimal model version.

[0069] When implementing dynamic weight allocation, a reinforcement learning environment (such as operating conditions) is defined, and a proxy model (such as Q-learning) adjusts the feature weights. For example, when wind speed is abnormal, the weight contribution of the wind speed feature is increased through the reward function. Those skilled in the art can implement this using the Reinforcement learning framework and update the weights periodically.

[0070] Example 2

[0071] like Figure 2 As shown, an FFU chemical filter lifecycle prediction system includes:

[0072] Data acquisition module: configured to acquire multi-dimensional dynamic parameters of the FFU chemical filter in real time via edge sensors;

[0073] Data processing module: configured to fill in missing values ​​and filter outliers in the collected data;

[0074] Feature engineering module: configured to construct temporally derived features and extract feature associations using Transformer encoding;

[0075] Model training module: Configured to use the Transformer-GBDT hybrid model for training, outputting remaining lifetime prediction results;

[0076] Optimization module: configured to optimize the hybrid model through hyperparameter search;

[0077] The system is applied in an industrial Internet of Things (IoT) environment, including at least one of edge computing devices or cloud servers.

[0078] It should be noted that by providing a modular system, the hardware deployment of the method in Example 1 can be realized, supporting industrial IoT environments (such as edge computing to reduce latency) and improving deployability (for example, in factory testing, the system supports equipment management at the level of thousands of people).

[0079] In practical implementation, the modular design involves a data acquisition module using a sensor network (such as a Raspberry Pi connected to a wind speed sensor); a data processing module running Python scripts on an edge device for interpolation and filtering; a feature engineering module using a server to process features; a model training module deployed on a cloud platform; and an optimization module that automatically performs parameter searches. This approach can be applied to the Industrial Internet of Things (IIoT): for example, in a chip factory, edge devices collect data, and cloud servers train models. Those skilled in the art can choose to deploy the modules using AWS or Azure platforms.

[0080] In this embodiment, a prediction output module is also included: configured to generate maintenance decisions and trigger equipment parameter adjustments based on the remaining life prediction results.

[0081] Through a closed-loop control system, maintenance decisions (such as replacing filters in advance) are generated, reducing manual intervention; parameter adjustments are triggered (such as adjusting wind speed to extend lifespan), improving system intelligence. In actual implementation, the prediction output module generates maintenance decisions (such as sending work order emails) based on model output results (such as remaining lifespan less than 10 days); and triggers equipment parameter adjustments (such as controlling FFU fan speed reduction via API). For example, when the predicted lifespan is short, the system automatically adjusts the wind speed to a safe value; those skilled in the art can implement the triggering mechanism using message queues (such as the MQTT protocol).

[0082] In this embodiment, the data acquisition module includes a compound concentration sensor and a wind speed sensor; the model training module is further configured to support online inference latency below a preset threshold.

[0083] In practical implementation, the data acquisition module is equipped with standard sensors (such as electrochemical compound concentration sensors to measure ppm values); the model training module optimizes the code (e.g., using Cython for acceleration) and sets a latency threshold (e.g., 100ms). For example, real-time inference on edge devices avoids network congestion; those skilled in the art can choose high-speed sensors and lightweight models (such as quantized Transformers) to implement this.

[0084] Finally, it should be noted that the technical essence of Embodiment 2 is the same as that of Embodiment 1. If there are any unclear technical aspects, please refer to Embodiment 1.

[0085] Example 3

[0086] A computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the method described in Embodiment 1.

[0087] In practice, Python programs are developed to implement methods S1-S5 and stored on storage media (such as USB or eMMC chips); the program is loaded when the processor executes it. For example, the user purchases the media, installs it on an industrial PC to run the prediction process, and those skilled in the art can package it into an executable file.

[0088] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

[0089] This approach can be easily implemented by those skilled in the art (such as data engineers or industrial maintenance personnel) without relying on complex code or tools. The key lies in the combination of data acquisition (sensor deployment), time-series processing (interpolation and filtering), feature engineering (computing derived features), model training (hybrid architecture), and optimization (adaptive mechanisms).

Claims

1. A method for predicting the life cycle of an FFU chemical filter, characterized in that, Includes the following steps: S1: Data Acquisition: Real-time acquisition of multi-dimensional dynamic parameters of the FFU chemical filter through edge sensors, including environmental parameters, operating parameters and performance indicators; S2: Data Processing: Imputing missing values ​​and filtering outliers in the collected data; S3: Feature Engineering: Based on the processed data, construct time-series derived features and use Transformer encoding to extract cross-period feature correlations; S4: Model Training: The extracted features are trained using a Transformer-GBDT hybrid model to output the remaining lifetime prediction results. The Transformer layer captures temporal dependencies, and the GBDT layer parses the feature interaction effects. S5: Model Optimization: Optimize the hybrid model through hyperparameter search; In S4, the Transformer-GBDT hybrid model includes: a Transformer encoding layer that encodes time-series data into a hidden state vector, a GBDT prediction layer that takes the hidden state vector and static features as input and outputs the remaining lifetime probability distribution by fitting the residuals; the GBDT prediction layer includes a gradient boosting tree variant, which is either XGBoost or LightGBM.

2. The method according to claim 1, characterized in that, In S1, the multi-dimensional dynamic parameters include filter type, zone layout, compound concentration, wind speed, rotation speed, downstream concentration, efficiency, consumable quantity, and usage time.

3. The method according to claim 1, characterized in that, In S2, the missing value imputation includes time-series correlation-based imputation, and the outlier filtering includes using a sliding window standard deviation combined with a dynamic threshold for removal.

4. The method according to claim 3, characterized in that, The missing value imputation further includes one of linear imputation or time series prediction imputation; the outlier filtering further includes one of clustering algorithm or deep learning anomaly detection.

5. The method according to claim 1, characterized in that, In S3, the temporal derived features include at least one of efficiency decay rate or concentration gradient; the Transformer encoding includes extracting feature associations using a Self-Attention mechanism.

6. The method according to claim 5, characterized in that, The Transformer encoding further includes one of a multi-head self-attention mechanism or a positional encoding variant.

7. The method according to claim 1, characterized in that, In S5, the hyperparameter search includes either Bayesian optimization or grid search; it also includes a dynamic weight allocation mechanism that adaptively adjusts the feature contribution based on changes in operating conditions. The dynamic weight allocation includes adjusting weights based on reinforcement learning.

8. A life cycle prediction system for an FFU chemical filter, characterized in that, include: Data acquisition module: configured to acquire multi-dimensional dynamic parameters of the FFU chemical filter in real time via edge sensors; Data processing module: configured to fill in missing values ​​and filter outliers in the collected data; Feature engineering module: configured to construct temporally derived features and extract feature associations using Transformer encoding; Model training module: Configured to use the Transformer-GBDT hybrid model for training, outputting remaining lifetime prediction results; The Transformer-GBDT hybrid model includes: a Transformer encoding layer that encodes time-series data into a hidden state vector; a GBDT prediction layer that takes the hidden state vector and static features as input and outputs the remaining lifetime probability distribution through residual fitting; the GBDT prediction layer includes a gradient boosting tree variant, which is either XGBoost or LightGBM. Optimization module: configured to optimize the hybrid model through hyperparameter search; The system is applied in an industrial Internet of Things (IoT) environment, including at least one of edge computing devices or cloud servers.

9. The system according to claim 8, characterized in that, It also includes a prediction output module: configured to generate maintenance decisions and trigger equipment parameter adjustments based on the remaining life prediction results.

Images