Intelligent optimization and quality prediction system for lubricating oil formula based on deep learning

Abstract

This invention discloses a deep learning-based intelligent optimization and quality prediction system for lubricating oil formulations, belonging to the field of intelligent manufacturing technology for lubricating oil formulations. It includes a data acquisition and preprocessing module for collecting multi-source heterogeneous data during the lubricating oil production process; a deep learning formulation optimization module that constructs a multi-objective optimization neural network model based on an improved Transformer architecture; an online quality prediction module employing a neural network model based on a temporal attention mechanism; and a feedback control module for comparing the predicted quality indicators output by the online quality prediction module with preset quality standards. This invention achieves efficient fusion and quality improvement of multi-source heterogeneous data, overcomes the modeling bottleneck of nonlinear interaction relationships among formulation components, realizes the generation of Pareto optimal solution sets for the collaborative optimization of multiple performance indicators, improves the foresight and reliability of quality prediction, and constructs a prediction-driven closed-loop adaptive control system.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology for lubricant formulations, and in particular to an intelligent optimization and quality prediction system for lubricant formulations based on deep learning. Background Technology

[0002] As an indispensable functional material in modern industry and transportation, the formulation design of lubricating oil directly determines the overall performance and service life of the product. Traditional lubricating oil formulation development mainly relies on the experience-based trial-and-error method of researchers, that is, screening combinations of base oils and additives through numerous experiments and repeatedly adjusting the ratio based on offline testing results. This method is not only time-consuming and costly, but also struggles to achieve an optimal balance among multiple performance indicators. Especially when facing increasingly stringent environmental regulations (such as API SP and ACEA C series standards) and diverse customized needs, the traditional method has shown significant limitations.

[0003] In recent years, with the development of artificial intelligence technology, some researchers have attempted to introduce machine learning into the field of lubricant formulation optimization. For example, response surface methodology (RSM) or support vector machine (SVM) is used to establish the mapping relationship between a single performance index and formulation components, and then grid search or genetic algorithms are used for formulation optimization. However, the above methods have the following shortcomings: First, traditional statistical models are difficult to capture the complex nonlinear interactions between base oil and various additives; second, existing technologies mostly optimize single performance indices and lack the ability to simultaneously optimize multiple objectives such as kinematic viscosity, viscosity index, flash point, pour point, acid value, and base value; third, the optimization process is disconnected from subsequent production quality control, failing to form a closed-loop technology system from formulation design to production execution.

[0004] In terms of quality prediction, existing lubricant production processes mainly rely on offline finished product testing, which has long testing cycles and significant time lags, making it impossible to detect quality deviations in a timely manner during production. While some online monitoring technologies can acquire real-time process parameters, they rely solely on simple linear regression or shallow neural networks for single-step predictions, making it difficult to effectively utilize long-range dependencies in time-series data, let alone perform multi-step ahead predictions of quality indicators for multiple key process stages. Furthermore, existing prediction models do not fully consider the constraints of physicochemical mechanisms, resulting in a lack of physical interpretability of prediction results under extreme operating conditions.

[0005] In process control, traditional PID control or manual adjustment based on experience are reactive compensation strategies. They intervene only after quality deviations actually occur or are confirmed by offline detection, and cannot adjust process parameters in advance based on predictive information. When raw material batches fluctuate or equipment operating conditions change, this lag in control often leads to the entire batch of products failing to meet quality standards, resulting in significant material waste and economic losses.

[0006] Therefore, there is an urgent need in this field for a technical solution for intelligent optimization and quality prediction of lubricant formulations that can organically integrate deep learning, multi-objective optimization, time-series prediction and feedback control, so as to achieve intelligent management and control of the entire process from formulation design to production execution. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a deep learning-based intelligent optimization and quality prediction system for lubricant formulations, effectively solving the problems in the background technology.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based intelligent optimization and quality prediction system for lubricant formulations, comprising: The data acquisition and preprocessing module is used to collect multi-source heterogeneous data in the lubricating oil production process, and to clean, normalize and perform feature engineering on the multi-source heterogeneous data to construct a standardized dataset. The multi-source heterogeneous data includes base oil physicochemical parameters, additive performance parameters, historical formula data, blending process parameters and finished product quality test data. The deep learning formulation optimization module constructs a multi-objective optimization neural network model based on an improved Transformer architecture. The model takes the standardized dataset as input and the mass fraction of each component in the lubricating oil formulation as output. It captures the nonlinear interaction relationship between formulation components through a self-attention mechanism and simultaneously optimizes multiple performance indicators of the lubricating oil by combining a multi-objective optimization algorithm, and outputs the optimal formulation scheme that meets the preset constraints. The online quality prediction module uses a neural network model based on a temporal attention mechanism. It takes the real-time process parameters in the current blending process and the optimal formulation as input to perform multi-step ahead prediction of the key quality indicators of the lubricating oil. The key quality indicators include kinematic viscosity, viscosity index, flash point, pour point, acid value and base value. The feedback control module is used to compare the predicted quality indicators output by the online quality prediction module with the preset quality standards. When the prediction deviation exceeds the preset threshold, it automatically generates process parameter adjustment instructions and sends them to the production execution system to realize closed-loop adaptive control of the production process.

[0009] Preferably, the data acquisition and preprocessing module includes: The multi-source data interface unit is used to interact with DCS, LIMS, and ERP systems via OPC-UA, MQTT, or Modbus protocols to collect data in real time on base oils, including kinematic viscosity at 40°C, kinematic viscosity at 100°C, viscosity index, flash point, pour point, density, and color parameters, as well as the active ingredient content, functional type, and recommended addition ratio parameters of additives. The data cleaning unit is used to identify and remove outlier data using an outlier detection method based on the isolated forest algorithm, and to process missing data using the KNN imputation method. The feature engineering unit is used to extract features and reduce dimensionality of the collected raw data, including: using Savitzky-Golay smoothing filter to eliminate baseline drift and noise interference in spectral data; using principal component analysis (PCA) to reduce the dimensionality of high-dimensional features and retain principal components with a cumulative variance contribution rate greater than 95%; and constructing interactive features, including the polarity matching index of base oil and additives, viscosity and temperature synergy coefficient, and antioxidant synergy coefficient.

[0010] Preferably, the deep learning recipe optimization module includes: The encoder unit adopts a multi-layer self-attention coding structure. The input layer receives the formula feature vector after feature engineering. The attention weight matrix between each formula component is calculated through a multi-head self-attention mechanism. The attention weight matrix represents the contribution of different base oil types, additive types and amounts to the overall performance of the lubricating oil. The multi-objective optimization decoder unit adopts a hybrid optimization strategy that combines the non-dominated sorting genetic algorithm (NSGA-III) based on the Pareto optimal front with neural networks. The optimization objectives are the kinematic viscosity of lubricating oil at 100℃, viscosity index, flash point, pour point, acid value, base value, sulfate ash content, and high-temperature high-shear viscosity at 150℃. The constraints are that the sum of the mass fraction of each component is 100%, the amount of each additive does not exceed its maximum allowable concentration, and the base oil type meets the API / ACEA standard. The unit outputs multiple non-dominated optimal formulation schemes located on the Pareto front. The formulation recommendation unit is used to select a single optimal formulation from the multiple non-dominated optimal formulation schemes according to the user's preset priority weights. The priority weights include cost priority, performance priority, or environmental priority. When the user does not specify a priority, the weight of each performance index is determined by default using an objective weighting strategy based on entropy weighting.

[0011] Preferably, the multi-head self-attention mechanism in the encoder unit adopts an improved ProbSparse self-attention mechanism, whose computational complexity is: , Denotes the asymptotic upper bound of the algorithm. The input size is the length of the input sequence. As the value approaches infinity, the rate of increase in algorithm execution time does not exceed [a certain percentage]. The growth rate, specifically includes: By measuring the Kullback-Leibler divergence between the query and the key, active queries with significantly different attention weight distributions are selected. Only the active queries are calculated with full self-attention, while the remaining queries are approximated using a mean aggregation strategy, thereby reducing computational overhead. The positional encoding information of the formulation components is introduced during the self-attention calculation process. The positional encoding adopts the method of adding the learnable sine and cosine positional encoding with the component type embedding vector, so that the model can distinguish the input order relationship of different component types.

[0012] Preferably, the online quality prediction module includes: The temporal feature extraction unit adopts an encoder and decoder structure based on the Informer architecture. The encoder receives the time series data of the current blending batch, including blending temperature, stirring speed, feed flow rate, blending time, and online viscosity detection value. It extracts multi-scale temporal features by compressing the sequence length layer by layer through self-attention distillation operation. The multi-step prediction unit adopts a generative decoder structure, takes the multi-scale temporal features output by the encoder as input, and directly generates the predicted values ​​of quality indicators for multiple future time steps through one forward propagation. The predicted time steps cover the key process stages from the mid-harmonization stage to the pre-finished product inspection stage. The uncertainty quantification unit is used to perform multiple forward inferences on the multi-step prediction unit using the Monte Carlo Dropout method, calculate the mean and variance of the predicted values ​​of each quality index, and trigger the data acquisition enhancement mechanism when the prediction variance exceeds the preset confidence interval, thereby increasing the sampling frequency of the batch reconciliation process.

[0013] Preferably, the multi-step prediction unit employs a progressive training strategy based on course learning during the training process, specifically including: In the first stage, the model is only trained to predict the quality indicators for the next time step, so that the model can grasp the short-term dynamic patterns. In the second stage, the prediction time step length is gradually increased, and constraint loss terms based on physical mechanisms are introduced. These constraint loss terms include: Arrhenius equation constraint on kinematic viscosity as a function of temperature, linear regression constraint on flash point and base oil fraction composition, and exponential decay constraint on pour point and pour point depressant addition amount. In the third stage, a complete multi-step prediction task is used for end-to-end fine-tuning, with the optimization objective being the weighted sum of the mean square error between the predicted and actual values ​​and the physical constraint loss term.

[0014] Preferably, the feedback control module includes: The deviation analysis unit is used to calculate the relative deviation between the predicted value and the target value of each key quality indicator. When the relative deviation of any indicator exceeds the corresponding preset threshold, it is determined that the batch has a quality risk. The preset threshold is determined according to the national standard tolerance range of each indicator. The preset threshold for kinematic viscosity is ±3%, the preset threshold for flash point is ±5℃, and the preset threshold for pour point is ±3℃. The parameter adjustment strategy generation unit is used to generate process parameter adjustment schemes based on a pre-trained reinforcement learning strategy network. The strategy network takes the current process state vector, the predicted deviation vector, and the adjustable parameter range as inputs and the adjusted process parameters as outputs. The adjustable parameters include the blending temperature setpoint, the stirring speed setpoint, and the additive addition amount. The safety constraint verification unit is used to verify the safety boundary of the process parameter adjustment scheme to ensure that the adjusted parameters do not exceed the safe operating range of the equipment. The safe operating range includes a mixing temperature not exceeding 180°C, a stirring speed not exceeding 90% of the rated speed of the equipment, and a single additive addition amount not exceeding 2% of the total formula.

[0015] Preferably, the reinforcement learning policy network is trained using the Proximal Policy (PPO) optimization algorithm, and its state space includes the current mixing temperature, stirring speed, real-time flow rate of each component, mixing time, and current online viscosity value; the action space includes temperature adjustment amount, stirring speed adjustment amount, and additive replenishment amount. The reward function is designed as follows: when all predicted quality indicators meet the standards, the reward value is positive and inversely proportional to the current formula cost; when there is a deviation in the prediction of quality indicators, the reward value is negative and directly proportional to the degree of deviation. At the same time, a penalty term is introduced to prevent process fluctuations caused by frequent adjustments.

[0016] This invention also provides a method for intelligent optimization and quality prediction of lubricant formulations based on deep learning, comprising the following steps: S1: Collect multi-source heterogeneous data from the lubricating oil production process through the data acquisition and preprocessing module, and perform data cleaning, normalization and feature engineering to build a standardized dataset; S2: Input the standardized dataset into the deep learning recipe optimization module, and output a set of Pareto optimal recipe schemes that meet the preset constraints based on the improved Transformer architecture multi-objective optimization neural network model; S3: Select a single optimal formula from the Pareto optimal formula set according to the user-preset priority weight, and send the optimal formula to the production execution system; S4: During the blending process, the online quality prediction module uses real-time process parameters as input to perform multi-step ahead prediction of key quality indicators; S5: The feedback control module compares the predicted quality indicators with the preset quality standards. When the prediction deviation exceeds the preset threshold, it automatically generates process parameter adjustment instructions and sends them to the production execution system to achieve closed-loop adaptive control.

[0017] The training process of the multi-objective optimization neural network model in step S2 includes: S21: Construct a training dataset, including historical recipe data and its corresponding quality inspection data, using the recipe data as input features and the quality inspection data as the optimization target label; S22: The input features are embedded using a pre-trained encoder. The pre-training uses a masked language model (MLM) strategy to randomly mask part of the recipe component information. The trained model recovers the masked information, enabling the model to learn the intrinsic relationship between the recipe components. S23: Input the embedded representation of the encoder output into the multi-objective optimization decoder, and use the NSGA-III algorithm and neural network joint optimization strategy to minimize the mean square error between the prediction quality index and the target value as the training objective, while satisfying the formulation component constraints. S24: Use an early stopping strategy to prevent model overfitting. When the multi-objective integrated loss on the validation set no longer decreases for 10 consecutive epochs, terminate training and save the optimal model parameters.

[0018] The beneficial effects of this invention are as follows: 1. This invention achieves efficient fusion and quality improvement of multi-source heterogeneous data: Through the multi-source data interface unit in the data acquisition and preprocessing module, the system supports multiple industrial communication protocols such as OPC-UA, MQTT, and Modbus, enabling seamless integration with DCS, LIMS, and ERP systems. This allows for the unified acquisition of basic oil physicochemical parameters, additive performance parameters, historical formulation data, blending process parameters, and finished product quality testing data. Combining outlier detection based on the isolated forest algorithm with KNN interpolation effectively solves the problems of missing data and noise interference in industrial settings. The feature engineering unit uses Savitzky-Golay smoothing filtering to eliminate spectral baseline drift, utilizes PCA dimensionality reduction to retain over 95% of the cumulative variance contribution rate, and constructs interactive features such as polarity matching index and viscosity-temperature synergy coefficient, significantly improving the input quality and training efficiency of subsequent models.

[0019] 2. This invention overcomes the modeling bottleneck of nonlinear interaction relationships among formulation components: The deep learning formulation optimization module adopts an improved Transformer architecture, calculating the attention weight matrix between each formulation component through a multi-head self-attention mechanism. This explicitly represents the contribution of different base oil types, additive types, and addition amounts to the overall performance of lubricating oil. In particular, the introduction of an improved ProbSparse self-attention mechanism, using Kullback-Leibler divergence to filter active queries, significantly reduces computational overhead while maintaining modeling accuracy, making it suitable for real-time optimization scenarios in high-dimensional formulation spaces. The combination of learnable positional encoding and component type embedding enables the model to distinguish the input order relationships of different component types, further enhancing the accuracy of formulation representation.

[0020] 3. This invention achieves Pareto optimal solution set generation through multi-performance index collaborative optimization: It employs a hybrid optimization strategy combining a non-dominated sorting genetic algorithm (NSGA-III) based on the Pareto optimal front with a neural network. Kinematic viscosity (100℃), viscosity index, flash point, pour point, acid value, base value, sulfate ash content, and high-temperature high-shear viscosity (150℃) are used as parallel optimization objectives. Simultaneously, constraints such as the sum of component mass fractions being 100%, maximum allowable additive concentration, and API / ACEA standard compliance are satisfied. The result is multiple non-dominated optimal formulation schemes located on the Pareto front. Compared to traditional single-objective optimization or weighted linear combination methods, this invention provides decision-makers with diverse trade-offs and enables flexible recommendations from a set of schemes to a single optimal solution through entropy weighting or user-defined priority weights, significantly improving the scientific rigor and adaptability of formulation design.

[0021] 4. This invention improves the foresight and reliability of quality prediction: The online quality prediction module adopts an encoder-decoder structure based on the Informer architecture. It compresses the time-series sequence length layer by layer through self-attention distillation, extracts multi-scale time-series features, and uses a generative decoder to directly output predicted quality indicators for multiple future time steps in a single forward propagation. Combined with the Monte Carlo Dropout uncertainty quantification method, the system can calculate the mean and variance of the predicted values. When the predicted variance exceeds the confidence interval, a data acquisition enhancement mechanism is automatically triggered, dynamically increasing the sampling frequency to improve prediction reliability. Furthermore, the progressive training strategy introduced in the course incorporates physical constraint loss terms such as the Arrhenius equation, flash point-fraction linear regression, and pour point-pour point depressant exponential decay, ensuring that the model prediction results possess both data-driven accuracy and consistency with physical mechanisms, effectively avoiding prediction distortion under extreme operating conditions.

[0022] 5. This invention constructs a prediction-driven closed-loop adaptive control system: the feedback control module compares multi-step ahead predicted quality indicators with preset national standards in real time. When the kinematic viscosity deviation exceeds ±3%, the flash point deviation exceeds ±5℃, and the pour point deviation exceeds ±3℃, etc., it automatically triggers a reinforcement learning strategy network based on the Proximal Policy Optimization (PPO) algorithm to generate process parameter adjustment schemes, including blending temperature setpoints, stirring speed setpoints, and additive replenishment amounts. The safety constraint verification unit ensures that the adjusted parameters do not exceed the safe operating range of the equipment (e.g., temperature not exceeding 180℃, speed not exceeding 90% of the rated value, and single additive replenishment amount not exceeding 2% of the total formula), realizing early intervention of quality risks and closed-loop adaptive control of the production process, significantly reducing the defect rate and raw material waste.

[0023] 6. This invention establishes a fully collaborative process encompassing formulation optimization, quality prediction, and feedback control: It deeply integrates three functional modules—formulation optimization, online prediction, and feedback control—changing the traditional model where formulation design and production control were separate. The pre-trained encoder uses a masked language model strategy to learn the intrinsic relationships between formulation components. Its embedded representation simultaneously serves both formulation optimization and quality prediction tasks, enabling knowledge transfer and sharing across different functional modules. The reward function of the reinforcement learning strategy network correlates quality achievement with formulation cost, automatically tending towards lower-cost solutions while ensuring quality, thus achieving a dual optimization of product quality and economic benefits. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the intelligent optimization and quality prediction system for lubricating oil formulations based on deep learning proposed in this invention. Figure 2 This is a schematic diagram of the architecture of the intelligent optimization and quality prediction system for lubricating oil formulation based on deep learning proposed in this invention. Figure 3 This is a schematic diagram illustrating the steps of the intelligent optimization and quality prediction method for lubricating oil formulation based on deep learning proposed in this invention.

[0025] In the diagram: 100, Data Acquisition and Preprocessing Module; 200, Deep Learning Recipe Optimization Module; 300, Online Quality Prediction Module; 400, Feedback Control Module. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0027] Reference Figure 1-2The intelligent optimization and quality prediction system for lubricating oil formulation based on deep learning includes a data acquisition and preprocessing module 100, a deep learning formulation optimization module 200, an online quality prediction module 300, and a feedback control module 400. Each module is deployed on the edge computing layer and cloud computing layer of the enterprise's intelligent manufacturing platform. The system's hardware environment includes: an industrial edge gateway deployed in the blending workshop (supporting OPC-UA Server, MQTT Broker, and Modbus TCP / RTU protocol conversion) for aggregating real-time process data from the DCS (Distributed Control System), quality inspection data from the LIMS (Laboratory Information Management System), and raw material batch information from the ERP system; a cloud-based GPU server cluster (configured with NVIDIA A100 Tensor Core GPUs) for carrying out model training and inference tasks for the deep learning formulation optimization module 200 and the online quality prediction module 300; and a workshop operator station equipped with a human-machine interface (HMI) for displaying the Pareto optimal formulation scheme set, real-time quality prediction curves, and process parameter adjustment commands.

[0028] The data acquisition and preprocessing module 100 is used to collect multi-source heterogeneous data in the lubricating oil production process, and to clean, normalize and perform feature engineering on the multi-source heterogeneous data to construct a standardized dataset. The multi-source heterogeneous data includes the physicochemical parameters of base oil, the performance parameters of additives, historical formula data, blending process parameters and finished product quality inspection data.

[0029] Furthermore, the data acquisition and preprocessing module 100 includes: The multi-source data interface unit is used to interact with DCS, LIMS, and ERP systems via OPC-UA, MQTT, or Modbus protocols to collect data in real time on base oils, including kinematic viscosity at 40°C, kinematic viscosity at 100°C, viscosity index, flash point, pour point, density, and color parameters, as well as the active ingredient content, functional type, and recommended addition ratio parameters of additives. The data cleaning unit is used to identify and remove outlier data using an outlier detection method based on the isolated forest algorithm, and to process missing data using the KNN imputation method. The feature engineering unit is used to extract features and reduce dimensionality of the collected raw data, including: using Savitzky-Golay smoothing filter to eliminate baseline drift and noise interference in spectral data; using principal component analysis (PCA) to reduce the dimensionality of high-dimensional features and retain principal components with a cumulative variance contribution rate greater than 95%; and constructing interactive features, including the polarity matching index of base oil and additives, viscosity and temperature synergy coefficient, and antioxidant synergy coefficient.

[0030] When implementing the data acquisition and preprocessing module 100: First, multi-source data interface and real-time acquisition: Taking a production batch of a certain brand of SN 5W-30 gasoline engine oil as an example, the multi-source data interface unit acquires real-time process parameters of the blending process from the DCS system via the OPC-UA protocol. The sampling frequency is set to 1Hz, specifically including: blending temperature (range 60℃-150℃), stirring speed (range 50rpm-300rpm), feed flow rate of each component (base oil, viscosity index improver, detergent dispersant, antioxidant and corrosion inhibitor, friction improver, pour point depressant, etc.), and online viscosity detection values ​​(kinematic viscosity at 40℃ and 100℃). Offline detection data of the finished product is obtained from the LIMS system via the MQTT protocol, including kinematic viscosity (100℃), viscosity index, flash point (open cup method), pour point, acid value, base value, sulfate ash content, and high-temperature high-shear viscosity (150℃). The batch attribute data of base oils and additives are obtained from the ERP system via the Modbus TCP protocol, including the kinematic viscosity at 40°C, kinematic viscosity at 100°C, viscosity index, flash point, pour point, density, and color (ASTM D1500) of the base oils, as well as the content of active ingredients, functional types (such as ZDDP, PIBSA, PMA, etc.) and supplier-recommended addition ratios of additives.

[0031] The second step is data cleaning and missing value handling: The data cleaning unit first performs outlier detection on the collected raw data. An Isolation Forest algorithm is used, with the number of trees (n_estimators=200), subsample ratio (subsample=0.8), and contamination ratio (contamination=0.05). For a batch of base oil with a kinematic viscosity at 40℃, if its outlier score is below the threshold of -0.6, it is identified as outlier data and removed. For missing data due to communication interruptions, KNN imputation (K=5, Euclidean distance as the distance metric, and Hamming distance for categorical features) is used for imputation. For example, if a batch is missing online viscosity data for 10 time points during blending, the system performs weighted imputation based on the viscosity values ​​of 5 neighboring normal batches at the same process stage, with the weights inversely proportional to the similarity of process parameters.

[0032] Step 3, Feature Engineering Processing: The feature engineering unit processes the cleaned data as follows: (1) Spectral data smoothing: For the base oil property data detected online using near-infrared spectroscopy (NIR), Savitzky-Golay smoothing filter is used, with window length set to window_length=15 and polynomial order=3, which effectively eliminates spectral baseline drift and high-frequency noise.

[0033] (2) Dimensionality reduction of high-dimensional features: Principal component analysis (PCA) was used to reduce the dimensionality of the formulation feature vector containing 50 original features (7 physicochemical indicators of base oil × 4 types of base oil + 6 attributes of additives × 8 types of additives + 6 process parameters). By calculating the eigenvalues ​​of the covariance matrix, the first 12 principal components were selected, so that the cumulative variance contribution rate reached 96.8%, reducing the feature dimension from 50 to 12, which significantly reduced the model complexity while retaining key information.

[0034] (3) Interaction Feature Construction: The system constructs the following three types of interaction features: Polarity Compatibility Index: Characterizes the polarity compatibility between base oil and detergent-dispersant, calculated using the following formula: ,in , This refers to the mass fraction of the components. , This is a polar solubility parameter; Viscosity-temperature synergy coefficient: reflects the synergistic effect between viscosity index improvers and base oil viscosity-temperature properties. ,in The viscosity index of each component, , The kinematic viscosity of the blended oil at 40℃ and 100℃; Antioxidant synergistic effect coefficient: calculated based on the oxidation induction period (OIT) of each antioxidant and its synergistic / antagonistic effect matrix. ,in The synergistic effect coefficient (obtained by fitting historical accelerated oxidation test data).

[0035] After the above processing, a standardized dataset is constructed, and each numerical feature is normalized using Z-score. ,in , These are the mean and standard deviation of the corresponding features in the training set, respectively.

[0036] Among them, the deep learning formula optimization module 200 constructs a multi-objective optimization neural network model based on the improved Transformer architecture. The model takes a standardized dataset as input and the mass fraction of each component in the lubricating oil formula as output. It captures the nonlinear interaction relationship between the formula components through a self-attention mechanism, and combines a multi-objective optimization algorithm to simultaneously optimize multiple performance indicators of the lubricating oil, and outputs the optimal formula scheme that meets the preset constraints.

[0037] Furthermore, the deep learning recipe optimization module 200 includes: The encoder unit adopts a multi-layer self-attention coding structure. The input layer receives the formula feature vector after feature engineering. The attention weight matrix between each formula component is calculated through a multi-head self-attention mechanism. The attention weight matrix represents the contribution of different base oil types, additive types and amounts to the overall performance of the lubricating oil. The multi-objective optimization decoder unit adopts a hybrid optimization strategy that combines the non-dominated sorting genetic algorithm (NSGA-III) based on the Pareto optimal front with neural networks. The optimization objectives are the kinematic viscosity of lubricating oil at 100℃, viscosity index, flash point, pour point, acid value, base value, sulfate ash content, and high-temperature high-shear viscosity at 150℃. The constraints are that the sum of the mass fraction of each component is 100%, the amount of each additive does not exceed its maximum allowable concentration, and the base oil type meets the API / ACEA standard. The unit outputs multiple non-dominated optimal formulation schemes located on the Pareto front. The formulation recommendation unit is used to select a single optimal formulation from multiple non-dominated optimal formulation schemes based on the user's preset priority weights. Priority weights include cost priority, performance priority, or environmental priority. When the user does not specify a priority, the weight of each performance indicator is determined by default using an objective weighting strategy based on entropy weighting.

[0038] The multi-head self-attention mechanism in the encoder unit adopts an improved ProbSparse self-attention mechanism, with a computational complexity of O(L.logL), where O represents the asymptotic upper bound of the algorithm, and L is the input size, i.e., the length of the input sequence. When the input size L approaches infinity, the growth rate of the algorithm's execution time does not exceed the growth rate of L.logL, specifically including: By measuring the Kullback-Leibler divergence between the query and the key, active queries with significantly different attention weight distributions are selected. Full self-attention is calculated only for active queries, while the remaining queries are approximated using a mean aggregation strategy, thereby reducing computational overhead. The self-attention calculation process incorporates positional encoding information of the formulation components. This positional encoding employs learnable sine and cosine positional codes added to the component type embedding vector, enabling the model to distinguish the input order relationships of different component types. This leads to the optimal formulation scheme that satisfies preset constraints.

[0039] In its implementation, the deep learning recipe optimization module 200 employs a 6-layer self-attention encoding structure in its encoder unit. Each layer contains a multi-head self-attention sub-layer and a feedforward neural network sub-layer. The input layer receives a 12-dimensional recipe feature vector after feature engineering, which is then extended to the model dimension through linear projection. .

[0040] The improved ProbSparse self-attention mechanism is implemented as follows: For the attention calculation of the Lth layer, let the query matrix... , where n is the quantity of the formulation components (including base oil and various additives). =64. Calculate each query. Kullback-Leibler divergence with bond distribution: Set the active query filtering ratio α=0.2, only for The top 20% of active queries calculate complete self-attention: ; The remaining 80% of queries use mean aggregation: .

[0041] This mechanism reduces the computational complexity of self-attention from... Down to .

[0042] Positional encoding employs a learnable sine / cosine positional encoding method, added to the component type embedding vector. Let the component type embedding matrix be:

[0043] in, For the number of predefined component categories (such as API Group I / II / III base oils, PMA pour point depressants, PIBSA detergents, etc.).

[0044] Sine position encoding: , , The final input is represented as: ; The multi-objective optimization decoder unit employs a hybrid optimization strategy combining the NSGA-III algorithm and a neural network. The neural network consists of a 3-layer fully connected network (256→128→64→8), with the output layer corresponding to the predicted values ​​of 8 quality indicators. The parameter settings for the NSGA-III algorithm are as follows: population size... Crossover probability Probability of mutation Number of reference points (3-layer target partitioning), maximum number of iterations .

[0045] Optimize the target vector ,in: Kinematic viscosity at 100℃ and target value (e.g., 9.3-12.5 mm)2 Normalization bias of / s); Normalized deviation of viscosity index from target value (≥120); The negative deviation (maximum) of the flash point from the target value (≥200℃); : Positive deviation (minimization) between pour point and target value (≤-35℃); Positive deviation of acid value from target value (≤2.0 mgKOH / g); The negative deviation of the alkalinity value from the target value (≥8.0 mgKOH / g); Positive deviation of sulfate ash content from the target value (≤1.0%); Normalized deviation of high-temperature high-shear viscosity (150℃) from target value (2.6-2.9 mPa·s).

[0046] The constraints include: : (Conservation of mass); : (Maximum permissible concentration of additives); : (Base oil API standard compliance); Through 500 generations of evolution, the algorithm converges to the Pareto optimal front, outputting 20-30 non-dominated optimal recipes.

[0047] When using the recipe recommendation unit: if the user does not specify priorities, the system defaults to using the entropy weight method to determine the weights of each performance indicator. The Pareto frontier solution set is used as the sample matrix.

[0048] First, normalization is performed to obtain Y, and then the information entropy of each indicator is calculated. Then the weight Then, a weighted composite score is calculated for each scheme, and the scheme with the highest score is selected as the single optimal formula scheme and issued.

[0049] If the user specifies a "cost priority" strategy, the Pareto solution set will be filtered to select the subset that meets all national standards for performance indicators and the solution with the lowest raw material cost will be selected. If "performance priority" is specified, the solution with the smallest overall deviation of each indicator will be selected. If "environmental protection priority" is specified, the solution with the lowest sulfate ash content and sulfur / phosphorus content will be selected first.

[0050] Model Training Process: The training dataset contains 3200 historical recipe data entries from the past 5 years and their corresponding quality inspection data. A Masked Language Model (MLM) strategy is used to pre-train the encoder: 15% of the recipe component information is randomly masked (replaced with [MASK] tags). The trained model recovers the type and quantity of the masked components based on context, enabling the model to learn the intrinsic relationships between the recipe components. Pre-training parameters: learning rate 1e-4, batch size 64, number of training epochs 100.

[0051] The fine-tuning phase employs joint optimization using NSGA-III and a neural network: the training objective is to minimize the mean squared error (MSE) between the predicted quality indicators and the true values, while the formulation component constraints are embedded into the loss function using the Lagrange multiplier method. The AdamW optimizer is used with an initial learning rate of 5e-5 and a weight decay of 0.01. An early stopping strategy is implemented: training is terminated and the optimal model parameters are saved when the multi-objective combined loss (the weighted average of the MSEs of each indicator) on the validation set (20% of the data split) no longer decreases for 10 consecutive epochs.

[0052] Among them, the online quality prediction module 300 adopts a neural network model based on the time-series attention mechanism. It takes the real-time process parameters and the optimal formula scheme in the current blending process as input to perform multi-step ahead prediction of the key quality indicators of the lubricating oil. The key quality indicators include kinematic viscosity, viscosity index, flash point, pour point, acid value and base value.

[0053] Furthermore, the online quality prediction module 300 includes: The temporal feature extraction unit adopts an encoder and decoder structure based on the Informer architecture. The encoder receives the time series data of the current blending batch, including blending temperature, stirring speed, feed flow rate, blending time, and online viscosity detection value. It extracts multi-scale temporal features by compressing the sequence length layer by layer through self-attention distillation operation. The multi-step prediction unit adopts a generative decoder structure, takes the multi-scale temporal features output by the encoder as input, and directly generates the predicted values ​​of quality indicators for multiple future time steps through one forward propagation. The predicted time steps cover the key process stages from the mid-harmonization stage to the pre-finished product inspection stage. The uncertainty quantification unit is used to perform multiple forward inferences on the multi-step prediction unit using the Monte Carlo Dropout method, calculate the mean and variance of the predicted values ​​of each quality index, and trigger the data acquisition enhancement mechanism when the prediction variance exceeds the preset confidence interval, thereby increasing the sampling frequency of the batch reconciliation process.

[0054] Furthermore, the multi-step prediction unit employs a progressive training strategy based on course learning during the training process, specifically including: In the first stage, the model is only trained to predict the quality indicators for the next time step, so that the model can grasp the short-term dynamic patterns. In the second stage, the prediction time step length is gradually increased, and constraint loss terms based on physical mechanisms are introduced. These constraint loss terms include: Arrhenius equation constraint on kinematic viscosity as a function of temperature, linear regression constraint on flash point and base oil fraction composition, and exponential decay constraint on pour point and pour point depressant addition amount. In the third stage, a complete multi-step prediction task is used for end-to-end fine-tuning, with the optimization objective being the weighted sum of the mean square error between the predicted and actual values ​​and the physical constraint loss term.

[0055] During the extraction of temporal features: Taking a harmonic batch (batch number: LUB2024061501) as an example, the time series feature extraction unit uses an Informer encoder to receive the time series data of this batch. Input sequence length... (Corresponding to 96 minutes, with a sampling interval of 1 minute), each time step has 6 feature dimensions: blending temperature, stirring speed, feed flow rate, blending time, online viscosity at 40℃, and online viscosity at 100℃.

[0056] The Informer encoder contains three self-attention distillation layers, each of which compresses the sequence length to half that of the previous layer through a self-attention distillation operation.

[0057] Level 1: Input length 96, output length 48; Level 2: Input length 48, output length 24; Level 3: Input length 24, output length 12.

[0058] The distillation process is as follows: ,in For the first The layer's self-attention output uses Conv1d as a 1D convolution (kernel size 3, stride 1) and MaxPool as a max pooling (kernel size 2, stride 2). The final extracted multi-scale temporal features have a dimension of 12×256.

[0059] The multi-step prediction unit adopts a generative decoder structure, taking the multi-scale temporal features output by the encoder as input, and directly generating the predicted quality indicators for the next 12 time steps (covering the critical stages from the mid-harmonic phase to the final product inspection) through a single forward propagation. The decoder structure consists of 2 self-attention layers + 3 fully connected layers (256→128→64→8), with the output layer corresponding to 8 quality indicators.

[0060] Predicted time step coverage: from the current time t to t+12min, corresponding to the heating stage, constant temperature stirring stage, homogenization stage and pre-finished product detection stage of the blending process.

[0061] Uncertainty quantification was performed using the Monte Carlo Dropout method: the Dropout layer was kept active during the inference phase (Dropout rate p=0.1), and T=100 forward inferences were performed. For the predicted kinematic viscosity at 100℃, the mean of the 100 inferences was calculated. With variance .

[0062] when (Preset confidence threshold, corresponding to a 95% confidence interval width exceeding 0.6 mm) 2 When the sampling frequency of the batch is 1Hz, the data acquisition enhancement mechanism is triggered, temporarily increasing the sampling frequency of the batch from 1Hz to 5Hz for 10 minutes, and the new data is appended to the input sequence for re-prediction.

[0063] The training of the multi-step prediction unit employs a three-stage curriculum learning strategy: Stage 1 (short-term prediction, Epochs 1-30): The model is trained only to predict the quality metrics for the next time step (t+1). The loss function is:

[0064] Learning rate 1e-3, batch size 32.

[0065] Phase Two (Asymptotic Multistep Prediction, Epochs 31-80): Gradually increase the prediction time step length, adding one time step every 5 epochs (increasing from 2 steps to 12 steps). Introduce a constraint loss term based on physical mechanisms:

[0066] Arrhenius equation constraints: ,in The predicted kinematic viscosity (100℃). The harmonization temperature corresponds to the time step. Pre-exponential factor, The activation energy (pre-calibrated via Arrhenius fitting); Flash point constraint: ,in The content of heavy fractions in base oil , These are the pre-calibrated linear regression coefficients; Pour point constraint: ,in Based on the pour point of the base oil, This refers to the amount of pour point depressant added. This is the attenuation coefficient.

[0067] Total loss function: Weight .

[0068] The third stage (end-to-end fine-tuning, Epochs 81-120): end-to-end training is performed using the complete 12-step prediction task, with the optimization objective being the weighted sum of the mean squared error between the predicted and actual values ​​and the physical constraint loss term. The learning rate is reduced to 2e-4, and cosine annealing scheduling is adopted.

[0069] The feedback control module 400 is used to compare the predicted quality indicators output by the online quality prediction module 300 with the preset quality standards. When the prediction deviation exceeds the preset threshold, it automatically generates process parameter adjustment instructions and sends them to the production execution system to realize closed-loop adaptive control of the production process.

[0070] Furthermore, the feedback control module 400 includes: The deviation analysis unit is used to calculate the relative deviation between the predicted value and the target value of each key quality indicator. When the relative deviation of any indicator exceeds the corresponding preset threshold, the batch is judged to have quality risk. The preset threshold is determined according to the national standard tolerance range of each indicator. The preset threshold for kinematic viscosity is ±3%, the preset threshold for flash point is ±5℃, and the preset threshold for pour point is ±3℃. The parameter adjustment strategy generation unit is used to generate process parameter adjustment schemes based on a pre-trained reinforcement learning strategy network. The strategy network takes the current process state vector, the predicted deviation vector, and the adjustable parameter range as inputs and the adjusted process parameters as outputs. The adjustable parameters include the blending temperature setpoint, the stirring speed setpoint, and the additive addition amount. The safety constraint verification unit is used to verify the safety boundaries of the process parameter adjustment scheme to ensure that the adjusted parameters do not exceed the safe operating range of the equipment. The safe operating range includes a mixing temperature not exceeding 180℃, a stirring speed not exceeding 90% of the rated speed of the equipment, and a single additive addition amount not exceeding 2% of the total formula.

[0071] Furthermore, the reinforcement learning policy network is trained using the proximal policy (PPO) optimization algorithm. Its state space includes the current mixing temperature, stirring speed, real-time flow rate of each component, mixing time, and current online viscosity value; the action space includes temperature adjustment amount, stirring speed adjustment amount, and additive replenishment amount. The reward function is designed as follows: when all predicted quality indicators meet the standards, the reward value is positive and inversely proportional to the current formula cost; when there is a deviation in the prediction of quality indicators, the reward value is negative and directly proportional to the degree of deviation. At the same time, a penalty term is introduced to prevent process fluctuations caused by frequent adjustments.

[0072] In the specific implementation of the feedback control module 400: First, deviation analysis: Taking a batch currently being blended as an example, the online quality prediction module outputs predicted quality indicators for the next 12 steps. The deviation analysis unit calculates the relative deviation between the predicted values ​​and target values ​​of each indicator. Target kinematic viscosity at 100℃: 10.5 mm 2 / s, predicted mean: 10.82 mm 2 / s, relative deviation: +3.05% (exceeding the ±3% threshold); Flash point target value: 215℃, predicted mean: 208℃, deviation: -7℃ (exceeding the ±5℃ threshold); Pour point target: -38℃, predicted mean: -36℃, deviation: +2℃ (not exceeding the ±3℃ threshold); The system determined that this batch had a quality risk and triggered the generation of a parameter adjustment strategy.

[0073] The second step is parameter adjustment strategy generation: The reinforcement learning policy network adopts the proximal policy optimization (PPO) algorithm. The network structure is: state encoder (3-layer LSTM, 256 hidden units) + policy head (2-layer fully connected, outputting the action mean and standard deviation) + value head (2-layer fully connected, outputting the state value).

[0074] state space ; Specifically, this includes: the current blending temperature. Current stirring speed Real-time flow rates of each component (Base oil: 850 kg / h, additives 1:12 kg / h, ...), blending time min, current online viscosity value Prediction deviation vectors for each quality indicator Action space Temperature adjustment amount Speed ​​adjustment amount Additive supplementation amount (Relative to the total amount of the formula).

[0075] Policy network output action distribution And select the value function The action that maximizes the outcome.

[0076] The reward function is designed as follows:

[0077] in, For reference formula cost, Based on the current formula cost, As the base for positive rewards, As a negative reward base, Assigning weights to each indicator. For the corresponding threshold, This represents the penalty coefficient for motion fluctuations.

[0078] In response to the deviation in this embodiment, the strategy network generates an adjustment plan: reducing the blending temperature from 135°C to 128°C (a decrease of 7°C), increasing the stirring speed from 180 rpm to 200 rpm (an increase of 20 rpm), and adding 0.5% pour point depressant (PPD). This plan is expected to reduce the kinematic viscosity deviation to +1.2% and the flash point deviation to -3°C within the next 10 minutes.

[0079] The third step involves the safety constraint verification unit performing boundary verification on the generated adjustment scheme. After adjustment, a temperature of 128℃ ≤ 180℃ (safe upper limit) is acceptable; After adjustment, the speed is 200 rpm ≤ rated speed 300 rpm × 90% = 270 rpm, which is considered passing. A single additive addition of 0.5% ≤ 2% of the total formula amount is acceptable. After the verification is successful, the system will send the process parameter adjustment command to the DCS system for execution via the OPC-UA protocol; if any condition is not met, the scheme will be rejected and a new action will be generated by resampling until all safety constraints are met.

[0080] Example 2, refer to Figure 3 This embodiment is an optimization based on Embodiment 1, specifically: a deep learning-based intelligent optimization and quality prediction method for lubricant formulations, including the following steps: S1: Collect multi-source heterogeneous data from the lubricating oil production process through the data acquisition and preprocessing module 100, and perform data cleaning, normalization and feature engineering to construct a standardized dataset; S2: Input the standardized dataset into the deep learning recipe optimization module 200, and output a set of Pareto optimal recipe schemes that meet the preset constraints based on the improved Transformer architecture multi-objective optimization neural network model. S3: Select a single optimal formula from the Pareto optimal formula set according to the user-preset priority weight, and send the optimal formula to the production execution system; S4: During the blending process, the online quality prediction module 300 uses real-time process parameters as input to perform multi-step ahead prediction of key quality indicators; S5: The feedback control module 400 compares the predicted quality indicators with the preset quality standards. When the prediction deviation exceeds the preset threshold, it automatically generates process parameter adjustment instructions and sends them to the production execution system to achieve closed-loop adaptive control.

[0081] Furthermore, the training process of the multi-objective optimization neural network model in step S2 includes: S21: Construct a training dataset, including historical recipe data and its corresponding quality inspection data, using the recipe data as input features and the quality inspection data as the optimization target label; S22: The input features are embedded using a pre-trained encoder. The pre-training uses a masked language model (MLM) strategy to randomly mask part of the recipe component information. The trained model recovers the masked information, enabling the model to learn the intrinsic relationship between the recipe components. S23: Input the embedded representation of the encoder output into the multi-objective optimization decoder, and use the NSGA-III algorithm and neural network joint optimization strategy to minimize the mean square error between the prediction quality index and the target value as the training objective, while satisfying the formulation component constraints. S24: Use an early stopping strategy to prevent model overfitting. When the multi-objective integrated loss on the validation set no longer decreases for 10 consecutive epochs, terminate training and save the optimal model parameters.

[0082] Specifically, taking the development of a new energy-saving gasoline engine oil (target specification: API SP5W-30) by a lubricant manufacturer as an example, this embodiment demonstrates the complete process of a deep learning-based intelligent optimization and quality prediction method for lubricant formulations: S1 Data Acquisition and Preprocessing: The system collected 3200 historical blending batch data points from the company over the past 5 years through multi-source data interfaces. This included batch attributes of four base oils (Group III base oil A / B, Group II+ base oil C, and synthetic ester base oil D) and eight functional additives (detergent, dispersant, antioxidant, anti-wear agent, friction modifier, viscosity index improver, pour point depressant, and antifoaming agent), as well as eight quality inspection data points for the corresponding finished products. After data cleaning and removal of 160 abnormal batches, KNN imputation was performed on the missing online viscosity data, ultimately constructing a standardized dataset containing 3040 samples. Feature engineering extracted 12-dimensional principal component features and three types of interaction features.

[0083] S2 Multi-Objective Optimization Recipe Generation: A standardized dataset was input into the deep learning formulation optimization module. The encoder learned the correlation representations between components through pre-training, and the multi-objective optimization decoder performed evolutionary optimization in an 8-dimensional objective space using the NSGA-III algorithm, with a population size of 200 and 500 generations. It ultimately converged to the Pareto front, obtaining 25 non-dominated optimal formulations with kinematic viscosities at 100°C ranging from 9.5 to 11.8 mm. 2 / s, viscosity index ranges from 125 to 142, flash point ranges from 208 to 228℃, and pour point ranges from -42 to -35℃.

[0084] S3: Single Optimal Solution Selection Since the user did not specify priorities, the system automatically calculated the weights of each indicator using the entropy weight method: kinematic viscosity 0.18, viscosity index 0.15, flash point 0.12, pour point 0.14, acid value 0.08, base value 0.10, sulfate ash content 0.13, and HTHS viscosity 0.10. Based on the weighted score, the scheme with the highest score from 25 options was selected as the single optimal formulation: Group III base oil A (72.5%), Group III base oil B (15.0%), synthetic ester base oil D (5.0%), detergent (3.2%), dispersant (2.0%), antioxidant (0.8%), anti-wear agent (0.9%), friction modifier (0.3%), viscosity index improver (0.2%), and pour point depressant (0.1%). This scheme meets all constraints and is estimated to reduce raw material costs by 8.3% compared to traditional empirical formulations.

[0085] S4: Online Multistep Quality Prediction During the blending process, the system acquires real-time process parameters at a frequency of 1 Hz. At the 60-minute mark of blending, the online quality prediction module generates multi-step predictions of quality indicators for the next 12 minutes (up to the 72nd minute) based on the time-series data from the previous 60 minutes. The prediction results show that the kinematic viscosity at 100°C will reach 10.9 mm² at the 68th minute. 2 / s (target 10.5±0.3), bias +3.8%, prediction variance 0.12, no data augmentation triggered.

[0086] S5: Closed-loop feedback control The deviation analysis unit determined that the kinematic viscosity deviation exceeded the threshold, triggering the PPO strategy network. Based on the current state, the strategy network generated adjustment instructions: reducing the blending temperature from 140℃ to 132℃ and increasing the stirring speed from 150 rpm to 175 rpm. After safety verification, the instructions were sent to the DCS system for execution at 61 minutes. Ten minutes later, a re-prediction showed that the kinematic viscosity deviation had decreased to +1.5%, and the quality risk was eliminated. The final offline test results showed a kinematic viscosity of 10.6 mm at 100℃. 2 / s, viscosity index 138, flash point 220℃, pour point -40℃, all indicators meet the requirements of API SP 5W-30 specifications, and the raw material cost is reduced by 8.3% compared with the traditional method, and the R&D cycle is shortened from the traditional 4-6 weeks to 72 hours.

[0087] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent optimization and quality prediction system for lubricating oil formulations, characterized in that, The system includes: The data acquisition and preprocessing module (100) is used to acquire multi-source heterogeneous data in the lubricating oil production process, and to clean, normalize and perform feature engineering on the multi-source heterogeneous data to construct a standardized dataset. The multi-source heterogeneous data includes the physicochemical parameters of base oil, the performance parameters of additives, historical formula data, blending process parameters and finished product quality test data. The deep learning formulation optimization module (200) constructs a multi-objective optimization neural network model based on the improved Transformer architecture. The model takes the standardized dataset as input and the mass fraction of each component in the lubricating oil formulation as output. It captures the nonlinear interaction relationship between the formulation components through a self-attention mechanism, and optimizes multiple performance indicators of the lubricating oil simultaneously by combining a multi-objective optimization algorithm, and outputs the optimal formulation scheme that meets the preset constraints. The online quality prediction module (300) adopts a neural network model based on the time-series attention mechanism. It takes the real-time process parameters in the current blending process and the optimal formula as input to perform multi-step ahead prediction of the key quality indicators of the lubricating oil. The key quality indicators include kinematic viscosity, viscosity index, flash point, pour point, acid value and base value. The feedback control module (400) is used to compare the predicted quality index output by the online quality prediction module (300) with the preset quality standard. When the prediction deviation exceeds the preset threshold, it automatically generates a process parameter adjustment instruction and sends it to the production execution system to realize closed-loop adaptive control of the production process.

2. The intelligent optimization and quality prediction system for lubricating oil formulation based on deep learning according to claim 1, characterized in that, The data acquisition and preprocessing module (100) includes: The multi-source data interface unit is used to interact with DCS, LIMS, and ERP systems via OPC-UA, MQTT, or Modbus protocols to collect data in real time on base oils, including kinematic viscosity at 40°C, kinematic viscosity at 100°C, viscosity index, flash point, pour point, density, and color parameters, as well as the active ingredient content, functional type, and recommended addition ratio parameters of additives. The data cleaning unit is used to identify and remove outlier data using an outlier detection method based on the isolated forest algorithm, and to process missing data using the KNN imputation method. The feature engineering unit is used to extract features and reduce dimensionality of the collected raw data, including: using Savitzky-Golay smoothing filter to eliminate baseline drift and noise interference in spectral data; using principal component analysis to reduce the dimensionality of high-dimensional features and retain principal components with a cumulative variance contribution rate greater than 95%; and constructing interactive features, including the polarity matching index of base oil and additives, viscosity and temperature synergy coefficient, and antioxidant synergy coefficient.

3. The intelligent optimization and quality prediction system for lubricating oil formulation based on deep learning according to claim 1, characterized in that, The deep learning recipe optimization module (200) includes: The encoder unit adopts a multi-layer self-attention coding structure. The input layer receives the formula feature vector after feature engineering. The attention weight matrix between each formula component is calculated through a multi-head self-attention mechanism. The attention weight matrix represents the contribution of different base oil types, additive types and amounts to the overall performance of the lubricating oil. The multi-objective optimization decoder unit adopts a hybrid optimization strategy that combines a non-dominated sorting genetic algorithm based on the Pareto optimal front with a neural network. The optimization objectives are the kinematic viscosity of the lubricating oil at 100℃, viscosity index, flash point, pour point, acid value, base value, sulfate ash content, and high-temperature high-shear viscosity at 150℃. The constraints are that the sum of the mass fractions of each component is 100%, the amount of each additive does not exceed its maximum allowable concentration, and the base oil type meets the API / ACEA standard. The unit outputs multiple non-dominated optimal formulation schemes located at the Pareto front. The formulation recommendation unit is used to select a single optimal formulation from the multiple non-dominated optimal formulation schemes according to the user's preset priority weights. The priority weights include cost priority, performance priority, or environmental priority. When the user does not specify a priority, the weight of each performance index is determined by default using an objective weighting strategy based on entropy weighting.

4. The intelligent optimization and quality prediction system for lubricating oil formulation based on deep learning according to claim 3, characterized in that, The multi-head self-attention mechanism in the encoder unit adopts an improved ProbSparse self-attention mechanism, specifically including: By measuring the Kullback-Leibler divergence between the query and the key, active queries with significantly different attention weight distributions are selected. Only the active queries are calculated with full self-attention, while the remaining queries are approximated using a mean aggregation strategy, thereby reducing computational overhead. The positional encoding information of the formulation components is introduced during the self-attention calculation process. The positional encoding adopts the method of adding the learnable sine and cosine positional encoding with the component type embedding vector, so that the model can distinguish the input order relationship of different component types.

5. The intelligent optimization and quality prediction system for lubricating oil formulation based on deep learning according to claim 1, characterized in that, The online quality prediction module (300) include: The temporal feature extraction unit adopts an encoder and decoder structure based on the Informer architecture. The encoder receives the time series data of the current blending batch, including blending temperature, stirring speed, feed flow rate, blending time, and online viscosity detection value. It extracts multi-scale temporal features by compressing the sequence length layer by layer through self-attention distillation operation. The multi-step prediction unit adopts a generative decoder structure, takes the multi-scale temporal features output by the encoder as input, and directly generates the predicted values ​​of quality indicators for multiple future time steps through one forward propagation. The predicted time steps cover the key process stages from the mid-harmonization stage to the pre-finished product inspection stage. The uncertainty quantification unit is used to perform multiple forward inferences on the multi-step prediction unit using the Monte Carlo Dropout method, calculate the mean and variance of the predicted values ​​of each quality index, and trigger the data acquisition enhancement mechanism when the prediction variance exceeds the preset confidence interval, thereby increasing the sampling frequency of the batch reconciliation process.

6. The intelligent optimization and quality prediction system for lubricating oil formulation based on deep learning according to claim 5, characterized in that, The multi-step prediction unit employs a progressive training strategy based on course learning during training, specifically including: In the first stage, the model is only trained to predict the quality indicators for the next time step, so that the model can grasp the short-term dynamic patterns. In the second stage, the prediction time step length is gradually increased, and constraint loss terms based on physical mechanisms are introduced. These constraint loss terms include: Arrhenius equation constraint on kinematic viscosity as a function of temperature, linear regression constraint on flash point and base oil fraction composition, and exponential decay constraint on pour point and pour point depressant addition amount. In the third stage, a complete multi-step prediction task is used for end-to-end fine-tuning, with the optimization objective being the weighted sum of the mean square error between the predicted and actual values ​​and the physical constraint loss term.

7. The intelligent optimization and quality prediction system for lubricating oil formulation based on deep learning according to claim 1, characterized in that, The feedback control module (400) includes: The deviation analysis unit is used to calculate the relative deviation between the predicted value and the target value of each key quality indicator. When the relative deviation of any indicator exceeds the corresponding preset threshold, it is determined that the batch has a quality risk. The preset threshold is determined according to the national standard tolerance range of each indicator. The preset threshold for kinematic viscosity is ±3%, the preset threshold for flash point is ±5℃, and the preset threshold for pour point is ±3℃. The parameter adjustment strategy generation unit is used to generate process parameter adjustment schemes based on a pre-trained reinforcement learning strategy network. The strategy network takes the current process state vector, the predicted deviation vector, and the adjustable parameter range as inputs and the adjusted process parameters as outputs. The adjustable parameters include the blending temperature setpoint, the stirring speed setpoint, and the additive addition amount. The safety constraint verification unit is used to verify the safety boundary of the process parameter adjustment scheme to ensure that the adjusted parameters do not exceed the safe operating range of the equipment. The safe operating range includes a mixing temperature not exceeding 180°C, a stirring speed not exceeding 90% of the rated speed of the equipment, and a single additive addition amount not exceeding 2% of the total formula.

8. The intelligent optimization and quality prediction system for lubricating oil formulation based on deep learning according to claim 7, characterized in that, The reinforcement learning strategy network is trained using a proximal policy optimization algorithm. Its state space includes the current mixing temperature, stirring speed, real-time flow rate of each component, mixing time, and current online viscosity value; the action space includes temperature adjustment amount, stirring speed adjustment amount, and additive replenishment amount. The reward function is designed as follows: when all predicted quality indicators meet the standards, the reward value is positive and inversely proportional to the current formula cost; when there is a deviation in the prediction of quality indicators, the reward value is negative and directly proportional to the degree of deviation. At the same time, a penalty term is introduced to prevent process fluctuations caused by frequent adjustments.

Images